[Fdo-trac] [fdo-trac] #1018: Fix up ODBC provider on Linux

FDO trac_fdo at osgeo.org
Wed Oct 21 01:26:11 PDT 2026


#1018: Fix up ODBC provider on Linux
---------------------------+----------------------
 Reporter:  jng            |        Owner:  jng
     Type:  defect         |       Status:  closed
 Priority:  major          |    Milestone:  4.2.0
Component:  ODBC Provider  |      Version:
 Severity:  3              |   Resolution:  fixed
 Keywords:                 |  External ID:
---------------------------+----------------------
Changes (by jng):

 * resolution:   => fixed
 * status:  assigned => closed

Comment:

 In [changeset:"8417" 8417]:
 {{{#!CommitTicketReference repository="" revision="8417"
 Port the ODBC provider to Linux and activate its SQL Server and Access
 suites

 The ODBC provider built against unixODBC but had never been exercised
 here: its
 unit-test fixtures were all behind #ifdef _WIN32, and the shared driver
 had a
 number of assumptions that either only held on Windows or only held where
 sizeof(long) == sizeof(int). Both provider suites now run on Linux --
 OdbcSqlServerTests against Microsoft's ODBC Driver 17 for SQL Server, and
 OdbcAccessTests against mdb-tools, the only file-based ODBC driver
 packaged by
 the distributions -- and both are green in Debug and Release.

 Defects that only show on LP64

   - define.c: RDBI_LONG is a C long to the RDBI layer, which sizes and
 strides a
     result array by sizeof(long) -- 8 here -- while get_o_t_s binds the
 4-byte
     SQL_C_LONG for it. The driver wrote one element per row where the
 layer read
     and stepped two, so every row from the second on delivered half of its
 own
     value and half of its neighbour's. Bind SQL_C_SBIGINT when the
 caller's
     buffer is wider than a SQLINTEGER.
   - desc_slct.c sized RDBI_INT by sizeof(long) for the same reason;
 RDBI_INT is
     an int to the layer, so size it by sizeof(int).
   - fetch.c stored SQL_ATTR_ROWS_FETCHED_PTR in an int. The driver writes
 a
     SQLULEN there, so the 8-byte store overran the adjacent stack slot.
   - geom.c declared the SQLParamData token as SQLSMALLINT, which truncates
 the
     pointer the driver hands back. It is a SQLPOINTER, as in the SQL
 Server
     Spatial provider's copy of this driver.
   - SchemaMgr/Ph/Mgr.cpp/.h: the ODBC reserved-word map was a class-
 static, and
     this file is compiled into both the provider DSO and the unit-test
     executable, so its destructor ran twice at process exit. Make it a
     function-local static, initialised and destroyed once.

 Driver handling off Windows

   - connect.c classified drivers by the Windows DLL names only, so the
 Linux
     SQL Server driver was never recognised; match msodbcsql and tdsodbc as
     substrings of SQL_DRIVER_NAME, which off Windows is the driver
 library's
     file name. SQLGetInfo(SQL_DRIVER_NAME) is not implemented by every
 driver
     either, so treat a failure as "unknown" and fall back to SQL_DBMS_NAME
     instead of failing the connection, and accept SQL_SUCCESS_WITH_INFO.
   - Odbc.cpp: RDBMS_LIB_NAME lost its ".so" suffix and getComDir() matches
 it as
     a prefix (the loaded name is versioned, libODBCProvider-4.2.0.so) and
     tolerates link-map entries with no slash, such as linux-vdso.so.1,
 which
     previously produced a NULL dereference.
   - FdoRdbmsOdbcConnectionInfo.cpp: recognise mdb-tools as a file
 provider, and
     accept the Database= keyword alongside DBQ=, which is the spelling the
     driver uses in a DSN.
   - FdoRdbmsOdbcConnection.cpp and FdoRdbmsFilterProcessor.cpp/.h: mdb-
 tools'
     SQL parser accepts neither a table alias nor a table-qualified column
 name,
     so FDO's "SELECT T.C FROM T ALIAS" was a parse error. Add a
 mUseTablePrefix
     switch beside mUseTableAliases and turn both off for connections that
     identify the mdb-tools driver. Every other driver keeps the SQL it
 had.

 mdb-tools, the Linux Access driver

   - constants.h and connect.c: add ODBCDriverType_MDBTools, classified
 from
     SQL_DBMS_NAME, which is the only identification the driver offers.
   - pkeys_.c: skip the SQLSpecialColumns approach for it. The driver
 returns
     success but its result set cannot be fetched -- it segfaults in libmdb
 --
     so identity properties come from the SQLColumns fallback.
   - objects_.c: drop the VIEW rows. mdb-tools lists Access saved queries
 as
     views but cannot execute them, and it ignores both the name and the
 type
     filter passed to SQLTables, so a view could not be described or
 selected.
   - col_.c: describe the probe query's columns with SQLDescribeCol for
 this
     driver. mdb-tools implements only SQL_DESC_NAME, SQL_DESC_TYPE,
     SQL_DESC_LENGTH, SQL_DESC_UNSIGNED and SQL_DESC_DISPLAY_SIZE -- every
 other
     SQLColAttribute is an error -- and its SQLColumns reports
 SQL_UNKNOWN_TYPE
     for many Access columns, while SQLDescribeCol reports the real type.
 The
     probe also stops sending a constant WHERE predicate, which the driver
     neither needs (the probe never fetches a row) nor releases.
   - desc_slct.c: execute the prepared statement before describing it. mdb-
 tools
     reports no result columns until SQLExecute, and FDO describes the
 select
     list first, so every query came back with zero columns.

 Test harness and test cases

   - OdbcConnectionUtil: name the SQL Server driver by
 FDO_SQLSERVER_ODBC_DRIVER
     (default ODBC Driver 17 for SQL Server) and the Access driver by
     FDO_ACCESS_ODBC_DRIVER (default MDBTools off Windows), and use them in
 the
     generated connection strings. Add GetAccessConnectString(), which
 builds
     DRIVER={...};DBQ=<test data>/MSTest.mdb -- mdb-tools needs no DSN and
 takes
     Database= only in a DSN -- and resolve the test data directory from
     /proc/self/exe, as the Windows version does with GetModuleFileName, so
 the
     path is sized for a Linux path rather than MAX_PATH.
 Connection_WithDSN
     falls back to the DRIVER= string off Windows, since Setup*DSN() is
     Windows-only, and do_rdbi_init() initialises the context on every
 platform
     instead of returning success without doing so.
   - OdbcBaseSetup: Sql2DbAutocommit() runs data-store DDL (CREATE/DROP
 DATABASE)
     with autocommit temporarily on, which SQL Server requires, and
 restores it
     before re-throwing.
   - Registrations and implementations move out of #ifdef _WIN32 for the
 SQL
     Server fixtures in OdbcFdoDeleteTest, OdbcFdoInsertTest,
 OdbcFdoUpdateTest,
     OdbcFdoSelectTest, OdbcFdoAdvancedSelectTest, OdbcFdoSchemaTest,
     OdbcFdoMultiThreadTest, OdbcFdoSqlCmdTest, OdbcDescribeSchemaTest,
     OdbcFdoConnectTest and OdbcFdoConnectionInfoTest, and for the read-
 only
     Access fixtures of OdbcDescribeSchemaTest, OdbcFdoSelectTest,
     OdbcFdoMultiThreadTest, OdbcFdoSchemaTest, OdbcFdoConnectTest and
     OdbcFdoConnectionInfoTest. What stays Windows-only is what needs a
 driver
     that does not exist here: the Excel, dBase and Text fixtures, and the
 Access
     write fixtures, the two view selects, the date-filter test and the
     computed-identifier arithmetic test, none of which mdb-tools can
 serve.
   - Fix the fixture constructors CppUnit instantiates: the private ones
 become
     public, since the suite registration constructs the fixture and GCC
 rejects
     a private constructor there. pOdbcConnectionUtil also loses its static
     linkage in OdbcTestRegister.cpp, so the other translation units can
 declare
     it extern.
   - SQLCMD_8BIT_TABLE_NAME is written as L"Fd\u00e9SqlTest": the source
 held a
     UTF-8 replacement character instead of the e-acute it meant, and SQL
 Server
     rejects the resulting name.

 Memory leaks (LeakSanitizer, Debug)

   - FdoConnectionInfoTest's mExpectedFileList is a vector of raw
     FileNameDataSet* that only ClearFileNameList() released, and the
 fixtures
     that populate it never call it, so every AddFileNameToList() left the
     FdoStringP-holding dataset behind (40 B per entry, direct). The
 destructor
     now releases whatever is still held.
   - The col_.c probe change above removed 62 of the 86 leaked objects the
     Access suite reported -- one glib chunk per WHERE expression,
 allocated by
     mdb-tools' own SQL engine and never freed -- leaving only the chunks
 its
     own filtered queries produce, which are the driver's defect rather
 than
     FDO's and are suppressed by the test harness rather than fixed here.

 Verification

   - OdbcSqlServerTests, ubuntu22 amd64, ODBC Driver 17 against the dbs
 compose
     mssql_cs instance: Debug OK (87), Release OK (87). 0 direct leak
 records;
     5,056 B in 4 indirect records, all from
     FdoExpressionEngineUtilDataReader::GetAggregateClassDef in the shared
     Expression Engine, reached through
     FdoAdvancedSelectTest::TestMaxBoolProperty.
   - OdbcAccessTests, ubuntu22 amd64, odbc-mdbtools 1.0.0: Debug OK (35),
     Release OK (35). The suite reports no records; before the probe change
 it
     reported 40,810 B in 168 allocations, including a direct record of
 26,144 B
     in 86 objects, of which 62 objects were the probe's WHERE expressions.
   - The other Odbc* suites are unchanged: MySQL and PostGIS run off their
 own
     init files and were not part of this work.

 API/ABI: the cycle is in freeze mode. Nothing under a shipped Inc/
 directory
 changes -- the headers touched here are under Providers/**/Src and are not
 installed -- so the provider-facing surface does not move and no baseline
 needs
 re-cutting.

 Fixes #1018
 }}}
-- 
Ticket URL: <https://trac.osgeo.org/fdo/ticket/1018#comment:1>
FDO <http://fdo.osgeo.org/>
Feature Data Objects


More information about the fdo-trac mailing list