Fix std::terminate() caused by uncaught exception in ORCA

src/backend/gpopt/CGPOptimizer.cpp

@@ -63,16 +63,17 @@ CGPOptimizer::GPOPTOptimizedPlan(
 	GPOS_CATCH_EX(ex)
 	{
 		// clone the error message before context free.
+		BOOL clone_failed = false;
 		CHAR *serialized_error_msg =
-			gpopt_context.CloneErrorMsg(MessageContext);
+			gpopt_context.CloneErrorMsg(MessageContext, &clone_failed);
 		// clean up context
 		gpopt_context.Free(gpopt_context.epinQuery, gpopt_context.epinPlStmt);
 
 		// Special handler for a few common user-facing errors. In particular,
 		// we want to use the correct error code for these, in case an application
 		// tries to do something smart with them.
 
-		if (GPOS_MATCH_EX(ex, gpdxl::ExmaGPDB, gpdxl::ExmiGPDBError))
+		if (clone_failed || GPOS_MATCH_EX(ex, gpdxl::ExmaGPDB, gpdxl::ExmiGPDBError))
 		{
 			PG_RE_THROW();
 		}

src/backend/gpopt/utils/COptTasks.cpp

@@ -162,13 +162,34 @@ SOptContext::Free(SOptContext::EPin input, SOptContext::EPin output) const
 //
 //---------------------------------------------------------------------------
 CHAR *
-SOptContext::CloneErrorMsg(MemoryContext context) const
+SOptContext::CloneErrorMsg(MemoryContext context, BOOL *clone_failed) const
 {
+	*clone_failed = false;
+
 	if (nullptr == context || nullptr == m_error_msg)
 	{
 		return nullptr;
 	}
-	return gpdb::MemCtxtStrdup(context, m_error_msg);
+
+	CHAR *error_msg;
+	GPOS_TRY
+	{
+#ifdef FAULT_INJECTOR
+		if (gpdb::InjectFaultInOptTasks("opt_clone_error_msg") == FaultInjectorTypeSkip)
+		{
+			GpdbEreport(ERRCODE_INTERNAL_ERROR, ERROR, "Injected error", nullptr);
+		}
+#endif
+		error_msg = gpdb::MemCtxtStrdup(context, m_error_msg);
+	}
+	GPOS_CATCH_EX(ex)
+	{
+		error_msg = nullptr;
+		*clone_failed = true;
+	}
+	GPOS_CATCH_END;
+
+	return error_msg;
 }
 
 

src/include/gpopt/utils/COptTasks.h

@@ -107,7 +107,7 @@ struct SOptContext
 	void Free(EPin input, EPin epinOutput) const;
 
 	// Clone the error message in given context.
-	CHAR *CloneErrorMsg(struct MemoryContextData *context) const;
+	CHAR *CloneErrorMsg(struct MemoryContextData *context, BOOL *clone_failed) const;
 
 	// casting function
 	static SOptContext *Cast(void *ptr);

src/test/regress/expected/gporca_faults.out

@@ -70,3 +70,28 @@ SELECT gp_inject_fault('opt_relcache_translator_catalog_access', 'reset', 1);
  Success:
 (1 row)
 
+-- Test to check that GPOPTOptimizedPlan() does not cause std::terminate() by throwing an uncaught exception.
+CREATE TABLE test_orca_uncaught_exc(a int, b int) DISTRIBUTED RANDOMLY;
+-- Since ORCA cannot optimize this query, an exception is generated. We then inject a second exception when the
+-- first is caught and verify that it is not propagated further (no std::terminate() is called and backend is alive).
+SELECT gp_inject_fault('opt_clone_error_msg', 'skip', 1);
+ gp_inject_fault 
+-----------------
+ Success:
+(1 row)
+
+SELECT sum(distinct a), count(distinct b) FROM test_orca_uncaught_exc;
+ sum | count 
+-----+-------
+     |     0
+(1 row)
+
+SELECT gp_inject_fault('opt_clone_error_msg', 'reset', 1);
+ gp_inject_fault 
+-----------------
+ Success:
+(1 row)
+
+-- start_ignore
+DROP TABLE test_orca_uncaught_exc;
+-- end_ignore

src/test/regress/expected/gporca_faults_optimizer.out

@@ -62,3 +62,24 @@ SELECT gp_inject_fault('opt_relcache_translator_catalog_access', 'reset', 1);
  Success:
 (1 row)
 
+-- Test to check that GPOPTOptimizedPlan() does not cause std::terminate() by throwing an uncaught exception.
+CREATE TABLE test_orca_uncaught_exc(a int, b int) DISTRIBUTED RANDOMLY;
+-- Since ORCA cannot optimize this query, an exception is generated. We then inject a second exception when the
+-- first is caught and verify that it is not propagated further (no std::terminate() is called and backend is alive).
+SELECT gp_inject_fault('opt_clone_error_msg', 'skip', 1);
+ gp_inject_fault 
+-----------------
+ Success:
+(1 row)
+
+SELECT sum(distinct a), count(distinct b) FROM test_orca_uncaught_exc;
+ERROR:  Injected error (COptTasks.cpp:180)
+SELECT gp_inject_fault('opt_clone_error_msg', 'reset', 1);
+ gp_inject_fault 
+-----------------
+ Success:
+(1 row)
+
+-- start_ignore
+DROP TABLE test_orca_uncaught_exc;
+-- end_ignore

src/test/regress/sql/gporca_faults.sql

@@ -30,3 +30,15 @@ SELECT * FROM func1_nosql_vol(5), foo;
 
 -- The fault should *not* be hit above when optimizer = off, to reset it now.
 SELECT gp_inject_fault('opt_relcache_translator_catalog_access', 'reset', 1);
+
+-- Test to check that GPOPTOptimizedPlan() does not cause std::terminate() by throwing an uncaught exception.
+CREATE TABLE test_orca_uncaught_exc(a int, b int) DISTRIBUTED RANDOMLY;
+-- Since ORCA cannot optimize this query, an exception is generated. We then inject a second exception when the
+-- first is caught and verify that it is not propagated further (no std::terminate() is called and backend is alive).
+SELECT gp_inject_fault('opt_clone_error_msg', 'skip', 1);
+SELECT sum(distinct a), count(distinct b) FROM test_orca_uncaught_exc;
+SELECT gp_inject_fault('opt_clone_error_msg', 'reset', 1);
+
+-- start_ignore
+DROP TABLE test_orca_uncaught_exc;
+-- end_ignore