Add support for coordinating record typmods among parallel workers.

Tuples can have type RECORDOID and a typmod number that identifies a blessed
TupleDesc in a backend-private cache.  To support the sharing of such tuples
through shared memory and temporary files, provide a typmod registry in
shared memory.

To achieve that, introduce per-session DSM segments, created on demand when a
backend first runs a parallel query.  The per-session DSM segment has a
table-of-contents just like the per-query DSM segment, and initially the
contents are a shared record typmod registry and a DSA area to provide the
space it needs to grow.

State relating to the current session is accessed via a Session object
reached through global variable CurrentSession that may require significant
redesign further down the road as we figure out what else needs to be shared
or remodelled.

Author: Thomas Munro
Reviewed-By: Andres Freund
Discussion: https://postgr.es/m/CAEepm=0ZtQ-SpsgCyzzYpsXS6e=kZWqk3g5Ygn3MDV7A8dabUA@mail.gmail.com

src/backend/access/common/Makefile

@@ -13,6 +13,6 @@ top_builddir = ../../../..
 include $(top_builddir)/src/Makefile.global
 
 OBJS = bufmask.o heaptuple.o indextuple.o printsimple.o printtup.o \
-	reloptions.o scankey.o tupconvert.o tupdesc.o
+	reloptions.o scankey.o session.o tupconvert.o tupdesc.o
 
 include $(top_srcdir)/src/backend/common.mk

src/backend/access/common/session.c

+/*-------------------------------------------------------------------------
+ *
+ * session.c
+ *		Encapsulation of user session.
+ *
+ * This is intended to contain data that needs to be shared between backends
+ * performing work for a client session.  In particular such a session is
+ * shared between the leader and worker processes for parallel queries.  At
+ * some later point it might also become useful infrastructure for separating
+ * backends from client connections, e.g. for the purpose of pooling.
+ *
+ * Currently this infrastructure is used to share:
+ * - typemod registry for ephemeral row-types, i.e. BlessTupleDesc etc.
+ *
+ * Portions Copyright (c) 2017, PostgreSQL Global Development Group
+ *
+ * src/backend/access/common/session.c
+ *
+ *-------------------------------------------------------------------------
+ */
+#include "postgres.h"
+
+#include "access/session.h"
+#include "storage/lwlock.h"
+#include "storage/shm_toc.h"
+#include "utils/memutils.h"
+#include "utils/typcache.h"
+
+/* Magic number for per-session DSM TOC. */
+#define SESSION_MAGIC						0xabb0fbc9
+
+/*
+ * We want to create a DSA area to store shared state that has the same
+ * lifetime as a session.  So far, it's only used to hold the shared record
+ * type registry.  We don't want it to have to create any DSM segments just
+ * yet in common cases, so we'll give it enough space to hold a very small
+ * SharedRecordTypmodRegistry.
+ */
+#define SESSION_DSA_SIZE					0x30000
+
+/*
+ * Magic numbers for state sharing in the per-session DSM area.
+ */
+#define SESSION_KEY_DSA						UINT64CONST(0xFFFFFFFFFFFF0001)
+#define SESSION_KEY_RECORD_TYPMOD_REGISTRY	UINT64CONST(0xFFFFFFFFFFFF0002)
+
+/* This backend's current session. */
+Session    *CurrentSession = NULL;
+
+/*
+ * Set up CurrentSession to point to an empty Session object.
+ */
+void
+InitializeSession(void)
+{
+	CurrentSession = MemoryContextAllocZero(TopMemoryContext, sizeof(Session));
+}
+
+/*
+ * Initialize the per-session DSM segment if it isn't already initialized, and
+ * return its handle so that worker processes can attach to it.
+ *
+ * Unlike the per-context DSM segment, this segement and its contents are
+ * reused for future parallel queries.
+ *
+ * Return DSM_HANDLE_INVALID if a segment can't be allocated due to lack of
+ * resources.
+ */
+dsm_handle
+GetSessionDsmHandle(void)
+{
+	shm_toc_estimator estimator;
+	shm_toc    *toc;
+	dsm_segment *seg;
+	size_t		typmod_registry_size;
+	size_t		size;
+	void	   *dsa_space;
+	void	   *typmod_registry_space;
+	dsa_area   *dsa;
+	MemoryContext old_context;
+
+	/*
+	 * If we have already created a session-scope DSM segment in this backend,
+	 * return its handle.  The same segment will be used for the rest of this
+	 * backend's lifetime.
+	 */
+	if (CurrentSession->segment != NULL)
+		return dsm_segment_handle(CurrentSession->segment);
+
+	/* Otherwise, prepare to set one up. */
+	old_context = MemoryContextSwitchTo(TopMemoryContext);
+	shm_toc_initialize_estimator(&estimator);
+
+	/* Estimate space for the per-session DSA area. */
+	shm_toc_estimate_keys(&estimator, 1);
+	shm_toc_estimate_chunk(&estimator, SESSION_DSA_SIZE);
+
+	/* Estimate space for the per-session record typmod registry. */
+	typmod_registry_size = SharedRecordTypmodRegistryEstimate();
+	shm_toc_estimate_keys(&estimator, 1);
+	shm_toc_estimate_chunk(&estimator, typmod_registry_size);
+
+	/* Set up segment and TOC. */
+	size = shm_toc_estimate(&estimator);
+	seg = dsm_create(size, DSM_CREATE_NULL_IF_MAXSEGMENTS);
+	if (seg == NULL)
+	{
+		MemoryContextSwitchTo(old_context);
+
+		return DSM_HANDLE_INVALID;
+	}
+	toc = shm_toc_create(SESSION_MAGIC,
+						 dsm_segment_address(seg),
+						 size);
+
+	/* Create per-session DSA area. */
+	dsa_space = shm_toc_allocate(toc, SESSION_DSA_SIZE);
+	dsa = dsa_create_in_place(dsa_space,
+							  SESSION_DSA_SIZE,
+							  LWTRANCHE_SESSION_DSA,
+							  seg);
+	shm_toc_insert(toc, SESSION_KEY_DSA, dsa_space);
+
+
+	/* Create session-scoped shared record typmod registry. */
+	typmod_registry_space = shm_toc_allocate(toc, typmod_registry_size);
+	SharedRecordTypmodRegistryInit((SharedRecordTypmodRegistry *)
+								   typmod_registry_space, seg, dsa);
+	shm_toc_insert(toc, SESSION_KEY_RECORD_TYPMOD_REGISTRY,
+				   typmod_registry_space);
+
+	/*
+	 * If we got this far, we can pin the shared memory so it stays mapped for
+	 * the rest of this backend's life.  If we don't make it this far, cleanup
+	 * callbacks for anything we installed above (ie currently
+	 * SharedRecordTypemodRegistry) will run when the DSM segment is detached
+	 * by CurrentResourceOwner so we aren't left with a broken CurrentSession.
+	 */
+	dsm_pin_mapping(seg);
+	dsa_pin_mapping(dsa);
+
+	/* Make segment and area available via CurrentSession. */
+	CurrentSession->segment = seg;
+	CurrentSession->area = dsa;
+
+	MemoryContextSwitchTo(old_context);
+
+	return dsm_segment_handle(seg);
+}
+
+/*
+ * Attach to a per-session DSM segment provided by a parallel leader.
+ */
+void
+AttachSession(dsm_handle handle)
+{
+	dsm_segment *seg;
+	shm_toc    *toc;
+	void	   *dsa_space;
+	void	   *typmod_registry_space;
+	dsa_area   *dsa;
+	MemoryContext old_context;
+
+	old_context = MemoryContextSwitchTo(TopMemoryContext);
+
+	/* Attach to the DSM segment. */
+	seg = dsm_attach(handle);
+	if (seg == NULL)
+		elog(ERROR, "could not attach to per-session DSM segment");
+	toc = shm_toc_attach(SESSION_MAGIC, dsm_segment_address(seg));
+
+	/* Attach to the DSA area. */
+	dsa_space = shm_toc_lookup(toc, SESSION_KEY_DSA, false);
+	dsa = dsa_attach_in_place(dsa_space, seg);
+
+	/* Make them available via the current session. */
+	CurrentSession->segment = seg;
+	CurrentSession->area = dsa;
+
+	/* Attach to the shared record typmod registry. */
+	typmod_registry_space =
+		shm_toc_lookup(toc, SESSION_KEY_RECORD_TYPMOD_REGISTRY, false);
+	SharedRecordTypmodRegistryAttach((SharedRecordTypmodRegistry *)
+									 typmod_registry_space);
+
+	/* Remain attached until end of backend or DetachSession(). */
+	dsm_pin_mapping(seg);
+	dsa_pin_mapping(dsa);
+
+	MemoryContextSwitchTo(old_context);
+}
+
+/*
+ * Detach from the current session DSM segment.  It's not strictly necessary
+ * to do this explicitly since we'll detach automatically at backend exit, but
+ * if we ever reuse parallel workers it will become important for workers to
+ * detach from one session before attaching to another.  Note that this runs
+ * detach hooks.
+ */
+void
+DetachSession(void)
+{
+	/* Runs detach hooks. */
+	dsm_detach(CurrentSession->segment);
+	CurrentSession->segment = NULL;
+	dsa_detach(CurrentSession->area);
+	CurrentSession->area = NULL;
+}

src/backend/access/common/tupdesc.c

@@ -184,6 +184,22 @@ CreateTupleDescCopyConstr(TupleDesc tupdesc)
 	return desc;
 }
 
+/*
+ * TupleDescCopy
+ *		Copy a tuple descriptor into caller-supplied memory.
+ *		The memory may be shared memory mapped at any address, and must
+ *		be sufficient to hold TupleDescSize(src) bytes.
+ *
+ * !!! Constraints and defaults are not copied !!!
+ */
+void
+TupleDescCopy(TupleDesc dst, TupleDesc src)
+{
+	memcpy(dst, src, TupleDescSize(src));
+	dst->constr = NULL;
+	dst->tdrefcount = -1;
+}
+
 /*
  * TupleDescCopyEntry
  *		This function copies a single attribute structure from one tuple

src/backend/access/transam/parallel.c

@@ -15,6 +15,7 @@
 #include "postgres.h"
 
 #include "access/parallel.h"
+#include "access/session.h"
 #include "access/xact.h"
 #include "access/xlog.h"
 #include "catalog/namespace.h"
@@ -36,6 +37,7 @@
 #include "utils/memutils.h"
 #include "utils/resowner.h"
 #include "utils/snapmgr.h"
+#include "utils/typcache.h"
 
 
 /*
@@ -51,8 +53,9 @@
 #define PARALLEL_MAGIC						0x50477c7c
 
 /*
- * Magic numbers for parallel state sharing.  Higher-level code should use
- * smaller values, leaving these very large ones for use by this module.
+ * Magic numbers for per-context parallel state sharing.  Higher-level code
+ * should use smaller values, leaving these very large ones for use by this
+ * module.
  */
 #define PARALLEL_KEY_FIXED					UINT64CONST(0xFFFFFFFFFFFF0001)
 #define PARALLEL_KEY_ERROR_QUEUE			UINT64CONST(0xFFFFFFFFFFFF0002)
@@ -63,6 +66,7 @@
 #define PARALLEL_KEY_ACTIVE_SNAPSHOT		UINT64CONST(0xFFFFFFFFFFFF0007)
 #define PARALLEL_KEY_TRANSACTION_STATE		UINT64CONST(0xFFFFFFFFFFFF0008)
 #define PARALLEL_KEY_ENTRYPOINT				UINT64CONST(0xFFFFFFFFFFFF0009)
+#define PARALLEL_KEY_SESSION_DSM			UINT64CONST(0xFFFFFFFFFFFF000A)
 
 /* Fixed-size parallel state. */
 typedef struct FixedParallelState
@@ -197,6 +201,7 @@ InitializeParallelDSM(ParallelContext *pcxt)
 	Size		segsize = 0;
 	int			i;
 	FixedParallelState *fps;
+	dsm_handle	session_dsm_handle = DSM_HANDLE_INVALID;
 	Snapshot	transaction_snapshot = GetTransactionSnapshot();
 	Snapshot	active_snapshot = GetActiveSnapshot();
 
@@ -211,6 +216,21 @@ InitializeParallelDSM(ParallelContext *pcxt)
 	 * Normally, the user will have requested at least one worker process, but
 	 * if by chance they have not, we can skip a bunch of things here.
 	 */
+	if (pcxt->nworkers > 0)
+	{
+		/* Get (or create) the per-session DSM segment's handle. */
+		session_dsm_handle = GetSessionDsmHandle();
+
+		/*
+		 * If we weren't able to create a per-session DSM segment, then we can
+		 * continue but we can't safely launch any workers because their
+		 * record typmods would be incompatible so they couldn't exchange
+		 * tuples.
+		 */
+		if (session_dsm_handle == DSM_HANDLE_INVALID)
+			pcxt->nworkers = 0;
+	}
+
 	if (pcxt->nworkers > 0)
 	{
 		/* Estimate space for various kinds of state sharing. */
@@ -226,8 +246,9 @@ InitializeParallelDSM(ParallelContext *pcxt)
 		shm_toc_estimate_chunk(&pcxt->estimator, asnaplen);
 		tstatelen = EstimateTransactionStateSpace();
 		shm_toc_estimate_chunk(&pcxt->estimator, tstatelen);
+		shm_toc_estimate_chunk(&pcxt->estimator, sizeof(dsm_handle));
 		/* If you add more chunks here, you probably need to add keys. */
-		shm_toc_estimate_keys(&pcxt->estimator, 6);
+		shm_toc_estimate_keys(&pcxt->estimator, 7);
 
 		/* Estimate space need for error queues. */
 		StaticAssertStmt(BUFFERALIGN(PARALLEL_ERROR_QUEUE_SIZE) ==
@@ -295,6 +316,7 @@ InitializeParallelDSM(ParallelContext *pcxt)
 		char	   *asnapspace;
 		char	   *tstatespace;
 		char	   *error_queue_space;
+		char	   *session_dsm_handle_space;
 		char	   *entrypointstate;
 		Size		lnamelen;
 
@@ -322,6 +344,13 @@ InitializeParallelDSM(ParallelContext *pcxt)
 		SerializeSnapshot(active_snapshot, asnapspace);
 		shm_toc_insert(pcxt->toc, PARALLEL_KEY_ACTIVE_SNAPSHOT, asnapspace);
 
+		/* Provide the handle for per-session segment. */
+		session_dsm_handle_space = shm_toc_allocate(pcxt->toc,
+													sizeof(dsm_handle));
+		*(dsm_handle *) session_dsm_handle_space = session_dsm_handle;
+		shm_toc_insert(pcxt->toc, PARALLEL_KEY_SESSION_DSM,
+					   session_dsm_handle_space);
+
 		/* Serialize transaction state. */
 		tstatespace = shm_toc_allocate(pcxt->toc, tstatelen);
 		SerializeTransactionState(tstatelen, tstatespace);
@@ -938,6 +967,7 @@ ParallelWorkerMain(Datum main_arg)
 	char	   *asnapspace;
 	char	   *tstatespace;
 	StringInfoData msgbuf;
+	char	   *session_dsm_handle_space;
 
 	/* Set flag to indicate that we're initializing a parallel worker. */
 	InitializingParallelWorker = true;
@@ -1064,6 +1094,11 @@ ParallelWorkerMain(Datum main_arg)
 	combocidspace = shm_toc_lookup(toc, PARALLEL_KEY_COMBO_CID, false);
 	RestoreComboCIDState(combocidspace);
 
+	/* Attach to the per-session DSM segment and contained objects. */
+	session_dsm_handle_space =
+		shm_toc_lookup(toc, PARALLEL_KEY_SESSION_DSM, false);
+	AttachSession(*(dsm_handle *) session_dsm_handle_space);
+
 	/* Restore transaction snapshot. */
 	tsnapspace = shm_toc_lookup(toc, PARALLEL_KEY_TRANSACTION_SNAPSHOT, false);
 	RestoreTransactionSnapshot(RestoreSnapshot(tsnapspace),
@@ -1110,6 +1145,9 @@ ParallelWorkerMain(Datum main_arg)
 	/* Shut down the parallel-worker transaction. */
 	EndParallelWorkerTransaction();
 
+	/* Detach from the per-session DSM segment. */
+	DetachSession();
+
 	/* Report success. */
 	pq_putmessage('X', NULL, 0);
 }

src/backend/storage/lmgr/lwlock.c

@@ -494,7 +494,7 @@ RegisterLWLockTranches(void)
 
 	if (LWLockTrancheArray == NULL)
 	{
-		LWLockTranchesAllocated = 64;
+		LWLockTranchesAllocated = 128;
 		LWLockTrancheArray = (char **)
 			MemoryContextAllocZero(TopMemoryContext,
 								   LWLockTranchesAllocated * sizeof(char *));
@@ -510,6 +510,12 @@ RegisterLWLockTranches(void)
 						  "predicate_lock_manager");
 	LWLockRegisterTranche(LWTRANCHE_PARALLEL_QUERY_DSA,
 						  "parallel_query_dsa");
+	LWLockRegisterTranche(LWTRANCHE_SESSION_DSA,
+						  "session_dsa");
+	LWLockRegisterTranche(LWTRANCHE_SESSION_RECORD_TABLE,
+						  "session_record_table");
+	LWLockRegisterTranche(LWTRANCHE_SESSION_TYPMOD_TABLE,
+						  "session_typmod_table");
 	LWLockRegisterTranche(LWTRANCHE_TBM, "tbm");
 
 	/* Register named tranches. */

src/backend/utils/cache/typcache.c

@@ -46,6 +46,8 @@
 #include "access/heapam.h"
 #include "access/htup_details.h"
 #include "access/nbtree.h"
+#include "access/parallel.h"
+#include "access/session.h"
 #include "catalog/indexing.h"
 #include "catalog/pg_am.h"
 #include "catalog/pg_constraint.h"
@@ -55,7 +57,9 @@
 #include "catalog/pg_type.h"
 #include "commands/defrem.h"
 #include "executor/executor.h"
+#include "lib/dshash.h"
 #include "optimizer/planner.h"
+#include "storage/lwlock.h"
 #include "utils/builtins.h"
 #include "utils/catcache.h"
 #include "utils/fmgroids.h"
@@ -142,6 +146,117 @@ typedef struct RecordCacheEntry
 	TupleDesc	tupdesc;
 } RecordCacheEntry;
 
+/*
+ * To deal with non-anonymous record types that are exchanged by backends
+ * involved in a parallel query, we also need a shared verion of the above.
+ */
+struct SharedRecordTypmodRegistry
+{
+	/* A hash table for finding a matching TupleDesc. */
+	dshash_table_handle record_table_handle;
+	/* A hash table for finding a TupleDesc by typmod. */
+	dshash_table_handle typmod_table_handle;
+	/* A source of new record typmod numbers. */
+	pg_atomic_uint32 next_typmod;
+};
+
+/*
+ * When using shared tuple descriptors as hash table keys we need a way to be
+ * able to search for an equal shared TupleDesc using a backend-local
+ * TupleDesc.  So we use this type which can hold either, and hash and compare
+ * functions that know how to handle both.
+ */
+typedef struct SharedRecordTableKey
+{
+	union
+	{
+		TupleDesc	local_tupdesc;
+		dsa_pointer shared_tupdesc;
+	};
+	bool		shared;
+} SharedRecordTableKey;
+
+/*
+ * The shared version of RecordCacheEntry.  This lets us look up a typmod
+ * using a TupleDesc which may be in local or shared memory.
+ */
+typedef struct SharedRecordTableEntry
+{
+	SharedRecordTableKey key;
+} SharedRecordTableEntry;
+
+/*
+ * An entry in SharedRecordTypmodRegistry's typmod table.  This lets us look
+ * up a TupleDesc in shared memory using a typmod.
+ */
+typedef struct SharedTypmodTableEntry
+{
+	uint32		typmod;
+	dsa_pointer shared_tupdesc;
+} SharedTypmodTableEntry;
+
+/*
+ * A comparator function for SharedTupleDescTableKey.
+ */
+static int
+shared_record_table_compare(const void *a, const void *b, size_t size,
+							void *arg)
+{
+	dsa_area   *area = (dsa_area *) arg;
+	SharedRecordTableKey *k1 = (SharedRecordTableKey *) a;
+	SharedRecordTableKey *k2 = (SharedRecordTableKey *) b;
+	TupleDesc	t1;
+	TupleDesc	t2;
+
+	if (k1->shared)
+		t1 = (TupleDesc) dsa_get_address(area, k1->shared_tupdesc);
+	else
+		t1 = k1->local_tupdesc;
+
+	if (k2->shared)
+		t2 = (TupleDesc) dsa_get_address(area, k2->shared_tupdesc);
+	else
+		t2 = k2->local_tupdesc;
+
+	return equalTupleDescs(t1, t2) ? 0 : 1;
+}
+
+/*
+ * A hash function for SharedRecordTableKey.
+ */
+static uint32
+shared_record_table_hash(const void *a, size_t size, void *arg)
+{
+	dsa_area   *area = (dsa_area *) arg;
+	SharedRecordTableKey *k = (SharedRecordTableKey *) a;
+	TupleDesc	t;
+
+	if (k->shared)
+		t = (TupleDesc) dsa_get_address(area, k->shared_tupdesc);
+	else
+		t = k->local_tupdesc;
+
+	return hashTupleDesc(t);
+}
+
+/* Parameters for SharedRecordTypmodRegistry's TupleDesc table. */
+static const dshash_parameters srtr_record_table_params = {
+	sizeof(SharedRecordTableKey),	/* unused */
+	sizeof(SharedRecordTableEntry),
+	shared_record_table_compare,
+	shared_record_table_hash,
+	LWTRANCHE_SESSION_RECORD_TABLE
+};
+
+/* Parameters for SharedRecordTypmodRegistry's typmod hash table. */
+static const dshash_parameters srtr_typmod_table_params = {
+	sizeof(uint32),
+	sizeof(SharedTypmodTableEntry),
+	dshash_memcmp,
+	dshash_memhash,
+	LWTRANCHE_SESSION_TYPMOD_TABLE
+};
+
 static HTAB *RecordCacheHash = NULL;
 
 static TupleDesc *RecordCacheArray = NULL;
@@ -168,6 +283,13 @@ static void TypeCacheConstrCallback(Datum arg, int cacheid, uint32 hashvalue);
 static void load_enum_cache_data(TypeCacheEntry *tcache);
 static EnumItem *find_enumitem(TypeCacheEnumData *enumdata, Oid arg);
 static int	enum_oid_cmp(const void *left, const void *right);
+static void shared_record_typmod_registry_detach(dsm_segment *segment,
+									 Datum datum);
+static void shared_record_typmod_registry_worker_detach(dsm_segment *segment,
+											Datum datum);
+static TupleDesc find_or_make_matching_shared_tupledesc(TupleDesc tupdesc);
+static dsa_pointer share_tupledesc(dsa_area *area, TupleDesc tupdesc,
+				uint32 typmod);
 
 
 /*
@@ -377,8 +499,8 @@ lookup_type_cache(Oid type_id, int flags)
 
 		/*
 		 * Reset info about hash functions whenever we pick up new info about
-		 * equality operator.  This is so we can ensure that the hash functions
-		 * match the operator.
+		 * equality operator.  This is so we can ensure that the hash
+		 * functions match the operator.
 		 */
 		typentry->flags &= ~(TCFLAGS_CHECKED_HASH_PROC);
 		typentry->flags &= ~(TCFLAGS_CHECKED_HASH_EXTENDED_PROC);
@@ -1243,6 +1365,33 @@ cache_record_field_properties(TypeCacheEntry *typentry)
 	typentry->flags |= TCFLAGS_CHECKED_FIELD_PROPERTIES;
 }
 
+/*
+ * Make sure that RecordCacheArray is large enough to store 'typmod'.
+ */
+static void
+ensure_record_cache_typmod_slot_exists(int32 typmod)
+{
+	if (RecordCacheArray == NULL)
+	{
+		RecordCacheArray = (TupleDesc *)
+			MemoryContextAllocZero(CacheMemoryContext, 64 * sizeof(TupleDesc));
+		RecordCacheArrayLen = 64;
+	}
+
+	if (typmod >= RecordCacheArrayLen)
+	{
+		int32		newlen = RecordCacheArrayLen * 2;
+
+		while (typmod >= newlen)
+			newlen *= 2;
+
+		RecordCacheArray = (TupleDesc *) repalloc(RecordCacheArray,
+												  newlen * sizeof(TupleDesc));
+		memset(RecordCacheArray + RecordCacheArrayLen, 0,
+			   (newlen - RecordCacheArrayLen) * sizeof(TupleDesc *));
+		RecordCacheArrayLen = newlen;
+	}
+}
 
 /*
  * lookup_rowtype_tupdesc_internal --- internal routine to lookup a rowtype
@@ -1273,16 +1422,54 @@ lookup_rowtype_tupdesc_internal(Oid type_id, int32 typmod, bool noError)
 		/*
 		 * It's a transient record type, so look in our record-type table.
 		 */
-		if (typmod < 0 || typmod >= NextRecordTypmod)
+		if (typmod >= 0)
+		{
+			/* It is already in our local cache? */
+			if (typmod < RecordCacheArrayLen &&
+				RecordCacheArray[typmod] != NULL)
+				return RecordCacheArray[typmod];
+
+			/* Are we attached to a shared record typmod registry? */
+			if (CurrentSession->shared_typmod_registry != NULL)
 			{
+				SharedTypmodTableEntry *entry;
+
+				/* Try to find it in the shared typmod index. */
+				entry = dshash_find(CurrentSession->shared_typmod_table,
+									&typmod, false);
+				if (entry != NULL)
+				{
+					TupleDesc	tupdesc;
+
+					tupdesc = (TupleDesc)
+						dsa_get_address(CurrentSession->area,
+										entry->shared_tupdesc);
+					Assert(typmod == tupdesc->tdtypmod);
+
+					/* We may need to extend the local RecordCacheArray. */
+					ensure_record_cache_typmod_slot_exists(typmod);
+
+					/*
+					 * Our local array can now point directly to the TupleDesc
+					 * in shared memory.
+					 */
+					RecordCacheArray[typmod] = tupdesc;
+					Assert(tupdesc->tdrefcount == -1);
+
+					dshash_release_lock(CurrentSession->shared_typmod_table,
+										entry);
+
+					return RecordCacheArray[typmod];
+				}
+			}
+		}
+
 		if (!noError)
 			ereport(ERROR,
 					(errcode(ERRCODE_WRONG_OBJECT_TYPE),
 					 errmsg("record type has not been registered")));
 		return NULL;
 	}
-		return RecordCacheArray[typmod];
-	}
 }
 
 /*
@@ -1303,7 +1490,7 @@ lookup_rowtype_tupdesc(Oid type_id, int32 typmod)
 	TupleDesc	tupDesc;
 
 	tupDesc = lookup_rowtype_tupdesc_internal(type_id, typmod, false);
-	IncrTupleDescRefCount(tupDesc);
+	PinTupleDesc(tupDesc);
 	return tupDesc;
 }
 
@@ -1321,7 +1508,7 @@ lookup_rowtype_tupdesc_noerror(Oid type_id, int32 typmod, bool noError)
 
 	tupDesc = lookup_rowtype_tupdesc_internal(type_id, typmod, noError);
 	if (tupDesc != NULL)
-		IncrTupleDescRefCount(tupDesc);
+		PinTupleDesc(tupDesc);
 	return tupDesc;
 }
 
@@ -1376,7 +1563,6 @@ assign_record_type_typmod(TupleDesc tupDesc)
 	RecordCacheEntry *recentry;
 	TupleDesc	entDesc;
 	bool		found;
-	int32		newtypmod;
 	MemoryContext oldcxt;
 
 	Assert(tupDesc->tdtypeid == RECORDOID);
@@ -1414,34 +1600,208 @@ assign_record_type_typmod(TupleDesc tupDesc)
 	recentry->tupdesc = NULL;
 	oldcxt = MemoryContextSwitchTo(CacheMemoryContext);
 
-	if (RecordCacheArray == NULL)
+	/* Look in the SharedRecordTypmodRegistry, if attached */
+	entDesc = find_or_make_matching_shared_tupledesc(tupDesc);
+	if (entDesc == NULL)
 	{
-		RecordCacheArray = (TupleDesc *) palloc(64 * sizeof(TupleDesc));
-		RecordCacheArrayLen = 64;
+		/* Reference-counted local cache only. */
+		entDesc = CreateTupleDescCopy(tupDesc);
+		entDesc->tdrefcount = 1;
+		entDesc->tdtypmod = NextRecordTypmod++;
 	}
-	else if (NextRecordTypmod >= RecordCacheArrayLen)
+	ensure_record_cache_typmod_slot_exists(entDesc->tdtypmod);
+	RecordCacheArray[entDesc->tdtypmod] = entDesc;
+	recentry->tupdesc = entDesc;
+
+	/* Update the caller's tuple descriptor. */
+	tupDesc->tdtypmod = entDesc->tdtypmod;
+
+	MemoryContextSwitchTo(oldcxt);
+}
+
+/*
+ * Return the amout of shmem required to hold a SharedRecordTypmodRegistry.
+ * This exists only to avoid exposing private innards of
+ * SharedRecordTypmodRegistry in a header.
+ */
+size_t
+SharedRecordTypmodRegistryEstimate(void)
+{
+	return sizeof(SharedRecordTypmodRegistry);
+}
+
+/*
+ * Initialize 'registry' in a pre-existing shared memory region, which must be
+ * maximally aligned and have space for SharedRecordTypmodRegistryEstimate()
+ * bytes.
+ *
+ * 'area' will be used to allocate shared memory space as required for the
+ * typemod registration.  The current process, expected to be a leader process
+ * in a parallel query, will be attached automatically and its current record
+ * types will be loaded into *registry.  While attached, all calls to
+ * assign_record_type_typmod will use the shared registry.  Worker backends
+ * will need to attach explicitly.
+ *
+ * Note that this function takes 'area' and 'segment' as arguments rather than
+ * accessing them via CurrentSession, because they aren't installed there
+ * until after this function runs.
+ */
+void
+SharedRecordTypmodRegistryInit(SharedRecordTypmodRegistry *registry,
+							   dsm_segment *segment,
+							   dsa_area *area)
+{
+	MemoryContext old_context;
+	dshash_table *record_table;
+	dshash_table *typmod_table;
+	int32		typmod;
+
+	Assert(!IsParallelWorker());
+
+	/* We can't already be attached to a shared registry. */
+	Assert(CurrentSession->shared_typmod_registry == NULL);
+	Assert(CurrentSession->shared_record_table == NULL);
+	Assert(CurrentSession->shared_typmod_table == NULL);
+
+	old_context = MemoryContextSwitchTo(TopMemoryContext);
+
+	/* Create the hash table of tuple descriptors indexed by themselves. */
+	record_table = dshash_create(area, &srtr_record_table_params, area);
+
+	/* Create the hash table of tuple descriptors indexed by typmod. */
+	typmod_table = dshash_create(area, &srtr_typmod_table_params, NULL);
+
+	MemoryContextSwitchTo(old_context);
+
+	/* Initialize the SharedRecordTypmodRegistry. */
+	registry->record_table_handle = dshash_get_hash_table_handle(record_table);
+	registry->typmod_table_handle = dshash_get_hash_table_handle(typmod_table);
+	pg_atomic_init_u32(&registry->next_typmod, NextRecordTypmod);
+
+	/*
+	 * Copy all entries from this backend's private registry into the shared
+	 * registry.
+	 */
+	for (typmod = 0; typmod < NextRecordTypmod; ++typmod)
 	{
-		int32		newlen = RecordCacheArrayLen * 2;
+		SharedTypmodTableEntry *typmod_table_entry;
+		SharedRecordTableEntry *record_table_entry;
+		SharedRecordTableKey record_table_key;
+		dsa_pointer shared_dp;
+		TupleDesc	tupdesc;
+		bool		found;
 
-		RecordCacheArray = (TupleDesc *) repalloc(RecordCacheArray,
-												  newlen * sizeof(TupleDesc));
-		RecordCacheArrayLen = newlen;
+		tupdesc = RecordCacheArray[typmod];
+		if (tupdesc == NULL)
+			continue;
+
+		/* Copy the TupleDesc into shared memory. */
+		shared_dp = share_tupledesc(area, tupdesc, typmod);
+
+		/* Insert into the typmod table. */
+		typmod_table_entry = dshash_find_or_insert(typmod_table,
+												   &tupdesc->tdtypmod,
+												   &found);
+		if (found)
+			elog(ERROR, "cannot create duplicate shared record typmod");
+		typmod_table_entry->typmod = tupdesc->tdtypmod;
+		typmod_table_entry->shared_tupdesc = shared_dp;
+		dshash_release_lock(typmod_table, typmod_table_entry);
+
+		/* Insert into the record table. */
+		record_table_key.shared = false;
+		record_table_key.local_tupdesc = tupdesc;
+		record_table_entry = dshash_find_or_insert(record_table,
+												   &record_table_key,
+												   &found);
+		if (!found)
+		{
+			record_table_entry->key.shared = true;
+			record_table_entry->key.shared_tupdesc = shared_dp;
+		}
+		dshash_release_lock(record_table, record_table_entry);
 	}
 
-	/* if fail in subrs, no damage except possibly some wasted memory... */
-	entDesc = CreateTupleDescCopy(tupDesc);
-	recentry->tupdesc = entDesc;
-	/* mark it as a reference-counted tupdesc */
-	entDesc->tdrefcount = 1;
-	/* now it's safe to advance NextRecordTypmod */
-	newtypmod = NextRecordTypmod++;
-	entDesc->tdtypmod = newtypmod;
-	RecordCacheArray[newtypmod] = entDesc;
+	/*
+	 * Set up the global state that will tell assign_record_type_typmod and
+	 * lookup_rowtype_tupdesc_internal about the shared registry.
+	 */
+	CurrentSession->shared_record_table = record_table;
+	CurrentSession->shared_typmod_table = typmod_table;
+	CurrentSession->shared_typmod_registry = registry;
 
-	/* report to caller as well */
-	tupDesc->tdtypmod = newtypmod;
+	/*
+	 * We install a detach hook in the leader, but only to handle cleanup on
+	 * failure during GetSessionDsmHandle().  Once GetSessionDsmHandle() pins
+	 * the memory, the leader process will use a shared registry until it
+	 * exits.
+	 */
+	on_dsm_detach(segment, shared_record_typmod_registry_detach, (Datum) 0);
+}
 
-	MemoryContextSwitchTo(oldcxt);
+/*
+ * Attach to 'registry', which must have been initialized already by another
+ * backend.  Future calls to assign_record_type_typmod and
+ * lookup_rowtype_tupdesc_internal will use the shared registry until the
+ * current session is detached.
+ */
+void
+SharedRecordTypmodRegistryAttach(SharedRecordTypmodRegistry *registry)
+{
+	MemoryContext old_context;
+	dshash_table *record_table;
+	dshash_table *typmod_table;
+
+	Assert(IsParallelWorker());
+
+	/* We can't already be attached to a shared registry. */
+	Assert(CurrentSession != NULL);
+	Assert(CurrentSession->segment != NULL);
+	Assert(CurrentSession->area != NULL);
+	Assert(CurrentSession->shared_typmod_registry == NULL);
+	Assert(CurrentSession->shared_record_table == NULL);
+	Assert(CurrentSession->shared_typmod_table == NULL);
+
+	/*
+	 * We can't already have typmods in our local cache, because they'd clash
+	 * with those imported by SharedRecordTypmodRegistryInit.  This should be
+	 * a freshly started parallel worker.  If we ever support worker
+	 * recycling, a worker would need to zap its local cache in between
+	 * servicing different queries, in order to be able to call this and
+	 * synchronize typmods with a new leader; see
+	 * shared_record_typmod_registry_detach().
+	 */
+	Assert(NextRecordTypmod == 0);
+
+	old_context = MemoryContextSwitchTo(TopMemoryContext);
+
+	/* Attach to the two hash tables. */
+	record_table = dshash_attach(CurrentSession->area,
+								 &srtr_record_table_params,
+								 registry->record_table_handle,
+								 CurrentSession->area);
+	typmod_table = dshash_attach(CurrentSession->area,
+								 &srtr_typmod_table_params,
+								 registry->typmod_table_handle,
+								 NULL);
+
+	MemoryContextSwitchTo(old_context);
+
+	/*
+	 * We install a different detach callback that performs a more complete
+	 * reset of backend local state.
+	 */
+	on_dsm_detach(CurrentSession->segment,
+				  shared_record_typmod_registry_worker_detach,
+				  PointerGetDatum(registry));
+
+	/*
+	 * Set up the session state that will tell assign_record_type_typmod and
+	 * lookup_rowtype_tupdesc_internal about the shared registry.
+	 */
+	CurrentSession->shared_typmod_registry = registry;
+	CurrentSession->shared_record_table = record_table;
+	CurrentSession->shared_typmod_table = typmod_table;
 }
 
 /*
@@ -1858,3 +2218,213 @@ enum_oid_cmp(const void *left, const void *right)
 	else
 		return 0;
 }
+
+/*
+ * Copy 'tupdesc' into newly allocated shared memory in 'area', set its typmod
+ * to the given value and return a dsa_pointer.
+ */
+static dsa_pointer
+share_tupledesc(dsa_area *area, TupleDesc tupdesc, uint32 typmod)
+{
+	dsa_pointer shared_dp;
+	TupleDesc	shared;
+
+	shared_dp = dsa_allocate(area, TupleDescSize(tupdesc));
+	shared = (TupleDesc) dsa_get_address(area, shared_dp);
+	TupleDescCopy(shared, tupdesc);
+	shared->tdtypmod = typmod;
+
+	return shared_dp;
+}
+
+/*
+ * If we are attached to a SharedRecordTypmodRegistry, use it to find or
+ * create a shared TupleDesc that matches 'tupdesc'.  Otherwise return NULL.
+ * Tuple descriptors returned by this function are not reference counted, and
+ * will exist at least as long as the current backend remained attached to the
+ * current session.
+ */
+static TupleDesc
+find_or_make_matching_shared_tupledesc(TupleDesc tupdesc)
+{
+	TupleDesc	result;
+	SharedRecordTableKey key;
+	SharedRecordTableEntry *record_table_entry;
+	SharedTypmodTableEntry *typmod_table_entry;
+	dsa_pointer shared_dp;
+	bool		found;
+	uint32		typmod;
+
+	/* If not even attached, nothing to do. */
+	if (CurrentSession->shared_typmod_registry == NULL)
+		return NULL;
+
+	/* Try to find a matching tuple descriptor in the record table. */
+	key.shared = false;
+	key.local_tupdesc = tupdesc;
+	record_table_entry = (SharedRecordTableEntry *)
+		dshash_find(CurrentSession->shared_record_table, &key, false);
+	if (record_table_entry)
+	{
+		Assert(record_table_entry->key.shared);
+		dshash_release_lock(CurrentSession->shared_record_table,
+							record_table_entry);
+		result = (TupleDesc)
+			dsa_get_address(CurrentSession->area,
+							record_table_entry->key.shared_tupdesc);
+		Assert(result->tdrefcount == -1);
+
+		return result;
+	}
+
+	/* Allocate a new typmod number.  This will be wasted if we error out. */
+	typmod = (int)
+		pg_atomic_fetch_add_u32(&CurrentSession->shared_typmod_registry->next_typmod,
+								1);
+
+	/* Copy the TupleDesc into shared memory. */
+	shared_dp = share_tupledesc(CurrentSession->area, tupdesc, typmod);
+
+	/*
+	 * Create an entry in the typmod table so that others will understand this
+	 * typmod number.
+	 */
+	PG_TRY();
+	{
+		typmod_table_entry = (SharedTypmodTableEntry *)
+			dshash_find_or_insert(CurrentSession->shared_typmod_table,
+								  &typmod, &found);
+		if (found)
+			elog(ERROR, "cannot create duplicate shared record typmod");
+	}
+	PG_CATCH();
+	{
+		dsa_free(CurrentSession->area, shared_dp);
+		PG_RE_THROW();
+	}
+	PG_END_TRY();
+	typmod_table_entry->typmod = typmod;
+	typmod_table_entry->shared_tupdesc = shared_dp;
+	dshash_release_lock(CurrentSession->shared_typmod_table,
+						typmod_table_entry);
+
+	/*
+	 * Finally create an entry in the record table so others with matching
+	 * tuple descriptors can reuse the typmod.
+	 */
+	record_table_entry = (SharedRecordTableEntry *)
+		dshash_find_or_insert(CurrentSession->shared_record_table, &key,
+							  &found);
+	if (found)
+	{
+		/*
+		 * Someone concurrently inserted a matching tuple descriptor since the
+		 * first time we checked.  Use that one instead.
+		 */
+		dshash_release_lock(CurrentSession->shared_record_table,
+							record_table_entry);
+
+		/* Might as well free up the space used by the one we created. */
+		found = dshash_delete_key(CurrentSession->shared_typmod_table,
+								  &typmod);
+		Assert(found);
+		dsa_free(CurrentSession->area, shared_dp);
+
+		/* Return the one we found. */
+		Assert(record_table_entry->key.shared);
+		result = (TupleDesc)
+			dsa_get_address(CurrentSession->area,
+							record_table_entry->key.shared);
+		Assert(result->tdrefcount == -1);
+
+		return result;
+	}
+
+	/* Store it and return it. */
+	record_table_entry->key.shared = true;
+	record_table_entry->key.shared_tupdesc = shared_dp;
+	dshash_release_lock(CurrentSession->shared_record_table,
+						record_table_entry);
+	result = (TupleDesc)
+		dsa_get_address(CurrentSession->area, shared_dp);
+	Assert(result->tdrefcount == -1);
+
+	return result;
+}
+
+/*
+ * Detach hook to forget about the current shared record typmod
+ * infrastructure.  This is registered directly in leader backends, and
+ * reached only in case of error or shutdown.  It's also reached indirectly
+ * via the worker detach callback below.
+ */
+static void
+shared_record_typmod_registry_detach(dsm_segment *segment, Datum datum)
+{
+	/* Be cautious here: maybe we didn't finish initializing. */
+	if (CurrentSession->shared_record_table != NULL)
+	{
+		dshash_detach(CurrentSession->shared_record_table);
+		CurrentSession->shared_record_table = NULL;
+	}
+	if (CurrentSession->shared_typmod_table != NULL)
+	{
+		dshash_detach(CurrentSession->shared_typmod_table);
+		CurrentSession->shared_typmod_table = NULL;
+	}
+	CurrentSession->shared_typmod_registry = NULL;
+}
+
+/*
+ * Deatch hook allowing workers to disconnect from shared record typmod
+ * registry.  The resulting state should allow a worker to attach to a
+ * different leader, if worker reuse pools are invented.
+ */
+static void
+shared_record_typmod_registry_worker_detach(dsm_segment *segment, Datum datum)
+{
+	/*
+	 * Forget everything we learned about record typmods as part of the
+	 * session we are disconnecting from, and return to the initial state.
+	 */
+	if (RecordCacheArray != NULL)
+	{
+		int32		i;
+
+		for (i = 0; i < RecordCacheArrayLen; ++i)
+		{
+			if (RecordCacheArray[i] != NULL)
+			{
+				TupleDesc	tupdesc = RecordCacheArray[i];
+
+				/*
+				 * Pointers to tuple descriptors in shared memory are not
+				 * reference counted, so we are not responsible for freeing
+				 * them.  They'll survive as long as the shared session
+				 * exists, which should be as long as the owning leader
+				 * backend exists.  In theory we do need to free local
+				 * reference counted tuple descriptors however, and we can't
+				 * do that with DescTupleDescRefCount() because we aren't
+				 * using a resource owner.  In practice we don't expect to
+				 * find any non-shared TupleDesc object in a worker.
+				 */
+				if (tupdesc->tdrefcount != -1)
+				{
+					Assert(tupdesc->tdrefcount > 0);
+					if (--tupdesc->tdrefcount == 0)
+						FreeTupleDesc(tupdesc);
+				}
+			}
+		}
+		pfree(RecordCacheArray);
+		RecordCacheArray = NULL;
+	}
+	if (RecordCacheHash != NULL)
+	{
+		hash_destroy(RecordCacheHash);
+		RecordCacheHash = NULL;
+	}
+	NextRecordTypmod = 0;
+	/* Call the code common to leader and worker detach. */
+	shared_record_typmod_registry_detach(segment, datum);
+}

src/backend/utils/init/postinit.c

@@ -21,6 +21,7 @@
 
 #include "access/heapam.h"
 #include "access/htup_details.h"
+#include "access/session.h"
 #include "access/sysattr.h"
 #include "access/xact.h"
 #include "access/xlog.h"
@@ -1027,6 +1028,9 @@ InitPostgres(const char *in_dbname, Oid dboid, const char *username,
 	/* initialize client encoding */
 	InitializeClientEncoding();
 
+	/* Initialize this backend's session state. */
+	InitializeSession();
+
 	/* report this backend in the PgBackendStatus array */
 	if (!bootstrap)
 		pgstat_bestart();

src/include/access/session.h

+/*-------------------------------------------------------------------------
+ *
+ * session.h
+ *	  Encapsulation of user session.
+ *
+ * Copyright (c) 2017, PostgreSQL Global Development Group
+ *
+ * src/include/access/session.h
+ *
+ *-------------------------------------------------------------------------
+ */
+#ifndef SESSION_H
+#define SESSION_H
+
+#include "lib/dshash.h"
+
+/* Defined in typcache.c */
+typedef struct SharedRecordTypmodRegistry SharedRecordTypmodRegistry;
+
+/*
+ * A struct encapsulating some elements of a user's session.  For now this
+ * manages state that applies to parallel query, but it principle it could
+ * include other things that are currently global variables.
+ */
+typedef struct Session
+{
+	dsm_segment *segment;		/* The session-scoped DSM segment. */
+	dsa_area   *area;			/* The session-scoped DSA area. */
+
+	/* State managed by typcache.c. */
+	SharedRecordTypmodRegistry *shared_typmod_registry;
+	dshash_table *shared_record_table;
+	dshash_table *shared_typmod_table;
+} Session;
+
+extern void InitializeSession(void);
+extern dsm_handle GetSessionDsmHandle(void);
+extern void AttachSession(dsm_handle handle);
+extern void DetachSession(void);
+
+/* The current session, or NULL for none. */
+extern Session *CurrentSession;
+
+#endif							/* SESSION_H */

src/include/access/tupdesc.h

@@ -92,6 +92,12 @@ extern TupleDesc CreateTupleDescCopy(TupleDesc tupdesc);
 
 extern TupleDesc CreateTupleDescCopyConstr(TupleDesc tupdesc);
 
+#define TupleDescSize(src) \
+	(offsetof(struct tupleDesc, attrs) + \
+	 (src)->natts * sizeof(FormData_pg_attribute))
+
+extern void TupleDescCopy(TupleDesc dst, TupleDesc src);
+
 extern void TupleDescCopyEntry(TupleDesc dst, AttrNumber dstAttno,
 				   TupleDesc src, AttrNumber srcAttno);
 

src/include/storage/lwlock.h

@@ -212,6 +212,9 @@ typedef enum BuiltinTrancheIds
 	LWTRANCHE_LOCK_MANAGER,
 	LWTRANCHE_PREDICATE_LOCK_MANAGER,
 	LWTRANCHE_PARALLEL_QUERY_DSA,
+	LWTRANCHE_SESSION_DSA,
+	LWTRANCHE_SESSION_RECORD_TABLE,
+	LWTRANCHE_SESSION_TYPMOD_TABLE,
 	LWTRANCHE_TBM,
 	LWTRANCHE_FIRST_USER_DEFINED
 }			BuiltinTrancheIds;

src/include/utils/typcache.h

@@ -18,6 +18,8 @@
 
 #include "access/tupdesc.h"
 #include "fmgr.h"
+#include "storage/dsm.h"
+#include "utils/dsa.h"
 
 
 /* DomainConstraintCache is an opaque struct known only within typcache.c */
@@ -143,6 +145,7 @@ typedef struct DomainConstraintRef
 	MemoryContextCallback callback; /* used to release refcount when done */
 } DomainConstraintRef;
 
+typedef struct SharedRecordTypmodRegistry SharedRecordTypmodRegistry;
 
 extern TypeCacheEntry *lookup_type_cache(Oid type_id, int flags);
 
@@ -164,4 +167,11 @@ extern void assign_record_type_typmod(TupleDesc tupDesc);
 
 extern int	compare_values_of_enum(TypeCacheEntry *tcache, Oid arg1, Oid arg2);
 
+extern size_t SharedRecordTypmodRegistryEstimate(void);
+
+extern void SharedRecordTypmodRegistryInit(SharedRecordTypmodRegistry *,
+							   dsm_segment *segment, dsa_area *area);
+
+extern void SharedRecordTypmodRegistryAttach(SharedRecordTypmodRegistry *);
+
 #endif							/* TYPCACHE_H */

src/tools/pgindent/typedefs.list

@@ -2016,6 +2016,10 @@ SharedInvalRelmapMsg
 SharedInvalSmgrMsg
 SharedInvalSnapshotMsg
 SharedInvalidationMessage
+SharedRecordTableKey
+SharedRecordTableEntry
+SharedRecordTypmodRegistry
+SharedTypmodTableEntry
 ShellTypeInfo
 ShippableCacheEntry
 ShippableCacheKey