psqlODBC release notes

psqlODBC 09.06.0200 Release

1. Start transactons first before calling SQLEndTran().
   
2. The failure of libpq_bind_exec() which returns NULL doesn't mean a FATAL error. Also make a check for disconnected connections.
   
3. Handle output parameters properly in build_libpq_bind_params.Also handle return values properly.
   
4. Make check of connection lost.
   
5. Revise recent lost connection check.
   
Check the lost connection

1. at the beginning of ODBC APIs.
   
2. at the beginning of FE-BE communications using extended protocol.
   
6. Stop specifying parameter datatype oids in parse message. Unfortunately it is inflexible and rather harmful. Though I don't know if it has much meaning, use cast instead just in order to keep regression test results. It's not necessary to cast any more the 2nd parameter of {call a_b_c_d_e(?, ?, ?, ?, ?)} in odbc-escapes-test.
   

psqlODBC 09.06.0100 Release

1. Optimization of a build system.
   
some adjustment and support Visual Studio 2015 for Windows.
2. Use SIZEOF_LONG_INT to check whether SQLROWSETSIZE exists.
   
SIZEOF_LONG_INT is defined in UnixODBC headers, while SIZEOF_LONG is defined by psqlodbc's ./configure. In odbcapi.c, we use it to determine what format modifier, %ld or %lld, to use for pointers. (Why don't we use %p?). We should rely on our autoconf check for that. In odbcapi.c, however, the check should match that in the ODBC headers, to determine whether SQLROWSIZE exists, so we should use SIZEOF_LONG_INT.
3. Fix the bug messages handle.
   
handle_error_message() and handle_notice_message() were unified into handle_pgres_error() in the wrong way. Unfortunately handle_error_message() never handled notice messages and set positive errornumbers which mean error_not_a_notice in psqlodbc. I renamed the constant CONNECTION_SERVER_REPORTED_ERROR to CONNECTION_SERVER_REPORTED_SEVERITY_FATAL and the constant CONNECTION_SERVER_REPORTED_WARNING to CONNECTION_SERVER_REPORTED_SEVERITY_ERROR so as to avoid confusion.
Reported by Shinji Mecha.
4. Allow setting isolation level before the establishment of connection.
   
5. Fixed return proper sqlstate error 08S01 in case of communication link failures.
   
Patch by Tsunakawa,Takayuki.
6. Check the existence of directory for logging.
   
7. Handle PG_DIAG_SEVERITY_NONLOCALIZED for servers whose version >= 9.6 so as to detect FATAL message properly.
   
8. Now default isolation level is server's default instead of fixed SQL_TXN_READ_COMMITTED.
   
Patch by Tsunakawa,Takayuki.
9. Make sure the isolation level is SQLUINTEGER and add the declaration of CC_get_isolation().
   
10. Change the default for ExtraSysTablePrefixes to an empty string.
    
11. Use lo_lseek64(lo_tell64) instead of lo_lseek(lo_tell) if the server's version >= 9.3. We are now able to handle > 2GB BLOBs.
    
12. Implement SQL_FECTH_BY_BOOKMARK operation of SQLBulkOperations properly.
    
13. Enable SQLBulkOperations in declare/fetch mode using long (>=12bytes) bookmarks.
    
Formerly the driver offered very simple bookmark support -- it is just the current row number. Now the driver offers more verbose bookmarks which contain KeySet informations (CTID (+ OID)). Though they consume 12bytes (row number + CTID) or 16bytes (row number + CTID + OID), they are useful in declare/fetch mode.
14. a hack to handle boolean items in VBA with MS Access.
    
VBA seems to transform the where condition a_boolean_item=True into ("a_boolean_item" = 1) which causes an ERROR:Operator does not exist boolean = integer . Here transforms it into ("a_boolean_item"='1') which seems safe in anyway.
15. Add a new option 'pqopt' which specifies libpq connection parameters as a conninfo style string.
    
The string must be enclosed with braces {} e.g. pqopt={sslcert=... sslkey=... sslrootcert=...} in the case of connection strings. Though host, port, dbname, user, password, sslmode, keepalives_idle or keepalive_interval parameters can be set using this option, they are ordinarily set by other options. When some settings for those parameters conflict with other ordinary options, connections are rejected.
16. Fix a corner case when parsing connection strings enclosed with braces.
    
17. As tuple_updated item in AddUpdated(), the parameter should be used. Deriving from other items caused bad results.
    
18. Change the default of ByteaAsLongVarBinary from 0 to 1.
    

psqlODBC 09.05.0400 Release

1. Enable COMMLOG and MYLOG button.
   
2. Default of *level of rollback on errors* is default(statement) not transaction.
   
3. Fix the bug SQLGetDescFieldW or SQLColAttributeW mistakenly returns SQL_ERROR.
   
4. Move src/*.exe to exe/*.exe on Windows.
   
Also use fc command instead of diff command on Windows. Fix a compilation error in large-object-data-at-exec-test.c regarding a type declaration in the middle of a code block.
Patch by Michael Paquier.
5. Ignore the difference between CR LF, LF and CR line break when comparing expected and result output.
   
6. Prior to Visual Studio 2015 Update 3, the hook functions of type PfnDliHook __pfnDliNotifyHook2 and __pfnDliFailureHook2 were non-const.
   
They were made const to improve security (global, writable function pointers are bad). If for backwards compatibility you require the hooks to be writable, define the macro DELAYIMP_INSECURE_WRITABLE_HOOKS prior to including this header and provide your own non-const definition of the hooks.
The following infomation is from Dave Cramer(davecramer@gmail.com).
7. Fix a regression test some failures.
   

psqlODBC 09.05.0300 Release

1. Largely reduce overhead of localtime() in SQLFetch
   
localtime() is a very costly system call causing serious grief on performance of the driver for OLTP-type applications using a lot SQLFetch and this one is being called for all data types when performing an in-line conversion, while it is being used only in the case where some of the date and/or timestamp fields are not initialized. While testing performance of an internal application, this patch has proved to improve the overall performance of an application by 20%.
2. Fix the use and the implementation of getNthValid().
   
3. Fix the handling of boundary_should_be_adjusted case when a rowset is cut in the middle of cached rows. Also fix the row info of SQL_ADDed tuples.
   
4. Estimate the rowsetSize including deleted tuples correctly so that rowset can be effectively handled.
   
5. Set once_reached_eof status when useDeclareFetch=0 so that QR_num_total_read() macro works well.
   
Don't forget to call CC_on_abort_partial() when rollback back to a savepoint. Now the driver passes cursor-block-delete test.
6. Refine SQLFetchScroll() for the orientations other than SQL_FETCH_NEXT/PRIOR.
   
Discard move_from_the_last stuff and call move_cursor_to_last() instead.
7. Refine cursors with KEYSET using oid or oid alternative as far as possible.
   
8. Preparation for handling inheritance.
   
9. Convert tableoid to relation name in case of inheritance.
   
10. fix buffer overrun in logging, and code cleanup.
    
Patch by Tsunakawa, Takayuki.
11. Implement SQL_QUERY_TIMEOUT.
    
Patch by Tsunakawa, Takayuki.
12. 'for' loop initial declarations are only allowed in C99 mode.
    

psqlODBC 09.05.0210 Release

1. Makefile.am in order to add new test case and expected output to distribution.
   
2. Use LoadLibraryEx for gxalib.dll load search path.
   

psqlODBC 09.05.0200 Release

1. Clear the parsed result before executing simple query so that SQLRowCount() returns.
   
This fix the bug reported by Walter Willmertinger.
2. Fix NULL dereferences which causes application crash.
   
Patch by Tsunakawa, Takayuki.
3. Don't initialize next_in_cache or rowstart_in_cache info when scrolling (block) cursors in declare/fetch mode.
   
This fixes the bug reported by John Kew.
4. Correct the detection of eof in useDeclare/fetch mode.
   
This change fixes the bug about fetching prior after scrolling next until eof reported by BenS.
5. Fix the bug about MONEY type.
   
6. Consider decimal_digits of SQL_TIME (and time field of PostgreSQL).
   
7. Don't forget to update num_total_read field of QResultClass record in case 'fetch 0 rows' after 'move'.
   
This fixes the bug 'When fetching a row "behind" the rowset by "fetch absolute" only the first ones of the result set can be fetched.' reported by Bernhard Goebel.
8. Make sure that EOF occurs when 'fetched < fetch requested'.
   
9. Correctly handle SQLGetData() when the target type is SQL_C_WCHAR and the buffer is small to get the whole data for some data types such as bool, date, time ... etc.
   
10. Refine SQLGetData().
    
1. Unify midtemp and moneytemp.
   
2. Variable wconverted is no longer needed.
   
3. Pre_convert for PG_TYPE_BOOL is not needed.
   
4. It's better to take the same road even in case of pre_convet.
   
11. Fixed possible once to build test suite in a different place other than the source code is, it's not now.
    
12. CHAR causes a compilation error in some environment. Use char instead of CHAR.
    
13. Correctly judge the SQLSTATE 01S06(Attempt to fetch before the result set returned the first rowset) for SQL_FETCH_PRIOR operations of SQFetchScroll.
    
14. Checking a flag bit mistakenly resets other flag bits.
    
Patch by Tsunakawa, Takayuki.
15. setup.c:test_connection() calls EN_Constructor() twice, leaking small memory.
    
Patch by tsunakawa, Takayuki.
16. The existing memory is not freed upon realloc() failure.
    
PGAPI_BindCol() doesn't check the failure of extend_getdata_info().
Patch by Tsunakawa, Takayuki.
17. The statement linked list is not properly chained, which throws away some statements.
    
Patch by Tsunakawa, Takayuki.
18. LINKMT is not used in the build process.
    
ADDL_INC should be ADD_INC. Just a typo. The CLEAN target doesn't delete .pdb files.
Patch by Tsunakawa, Takayuki.
19. Fix a parameter unmatch of snprintf.
    
20. In Windows the use of external variables timezone, tzname and daylight are deprecated. Use _timezone, _tzname and _daylight instead.
    
21. SQLS(G)etDescRecW instead of SQLS(G)etDescRec for Unicode driver.
    
22. The third stricmp() in writeDriverCommoninfo() is useless, because the second one causes the function to return.
    
1. SC_get_errormsg() definition is repeated twice in succession.
   
2. In info.c:SQLGetInfo(), "len = 0" is unnecessary because len is assigned at the end of the function.
   
3. In execute.c:PGAPI_ExecDirect(), flag doesn't need to be set as its value doesn't change.
   
4. PODBC_PER_STATEMENT_ROLLBACK is not used.
   
Patch by Tsunakawa, Takayuki.

psqlODBC 09.05.0100 Release

1. Use libpq for all communication with the server
   
Previously, libpq was only used for authentication. Using it for all communication lets us remove a lot of duplicated code. libpq is now required for building or using libpq.
2. Remove support for protocol versions older than 3.
   
Protocol version 3 was introduced in PostgreSQL 7.4, so this means that we no longer support server versions 7.3 or older. This simplified the code, making maintenance easier.
3. Remove support for ODBC versions older than 3.51.
   
The official binaries have been compiled in 3.51 mode for a very long time, so this has no effect on the official binaries.
4. Avoid one round-trip to the server when establishing connection.
   
When connecting to a server, the driver used to issue a query to determine the client_encoding setting being used, but that was unnecessary because all supported server versions send that information as part of the connection handshake when a connection is established. Remove the extra query, which should make connecting over a high-latency link faster.
5. Report max length of 'name' fields correctly.
   
PostgreSQL columns of type 'name', used e.g. in catalog tables to hold names of relations and other objects, were reported as having type VARCHAR(64). However, the real max length of a name field is only 63 bytes.
6. Remove 'Optimizer' configuration option
   
You can use the generic ConnSettings option with value "set geqo=off", to disable GEQO.
7. Fix title in 32-bit Unicode drivers setup dialog
   
The title on the 32-bit Unicode drivers's setup dialog on Windows incorrectly said "ANSI".
8. Don't export private symbols that are not part of the ODBC interface
   
This was causing problems when the driver was loaded together with another dynamically loaded library with a function of the same name. This was reported to happen when loading the odbc_fdw foreign data wrapper into a PostgreSQL server, because both the PostgreSQL server and the ODBC driver contain a function called check_client_encoding().
9. Report a sensible SQL_DESC_OCTET_LENGTH value for result set columns of type 'unknown'
   
If a query returns a column of type 'unknown', as happens e.g. in a query like "select 'foobar' from table" where the datatype of the constant 'foobar' is not specified, the SQL_DESC_OCTET_LENGTH property of the column was returned as invalid. Report it the same as 'varchar' field with no explicit length, instead.
10. Fix quoting bugs in sending integer query parameters to server
    
The drivers used to assume that if a parameter's SQL type is SQL_INTEGER or SQL_SMALLINT, the value does not require quoting when its send to the server. For example, "SELECT ?", with parameter 123 was translated to "SELECT 123", when UseServerSidePrepare was not enabled. However, there was no check that the query parameter in fact contained a valid integer, when replacing the parameter markers with their values. Also, in a query like "SELECT 0-?", a negative value needs to have parens around it, as in "SELECT 0-(-123)".
11. Don't reset autocommit when a connection is established
    
If autocommit is disabled on a connection, by calling SQLSetConnectAttr(SQL_ATTR_AUTOCOMMIT, SQL_AUTOCOMMIT_OFF, 0), before connecting with SQLDriverConnect(), autocommit was incorrectly reset back to on when the connection was established.
12. Send datatype information for query parameters, when known
    
If a query parameter is bound with a specific SQL type, pass on that information to the server. This makes the behaviour of queries like "SELECT '555' > ?" more sensible, where the result depends on whether the query parameter is interpreted as an integer or a string.
13. Use libpq defaults for Server, Username and Database settings
    
Server, Username and Database settings can now be left empty, in which case the corresponding libpq defaults will be used. The libpq default is to connect to a local server over Unix domain sockets (on Unix systems) or to localhost over TCP/IP (on Windows), if Server is left empty. The default username is the current operating system user, and the database to connect to is a database with the same name as the user. These built-in defaults can be changed by environment variables, however. This is particularly useful for connecting to a local installation via Unix domain sockets, as that is the default when Server is left empty, and was difficult to configure otherwise.
14. Fix bug in parsing E'' literals
    
The driver needs to parse a query string to determine whether a '?' character is a parameter marker, or if it's inside a quoted string or a SQL comment. This parsing did not correctly handle E''-style strings.
15. Remove 256 byte limit on the length of NOTICE messages
    
There was a built-in limit of 256 bytes on the length of NOTICE messages received from the server. Strings larger than 256 bytes were not returned to the application. That limit has been removed.
16. Fix parsing of strings containing escaped quotes with Parse=1
    
This fixes issues in extracting metadata like result set column names, with the Parse=1 option.
17. Fix connecting with a user or database name that contains spaces or quotes
    
User and database names were not quoted correctly when building a connection string. A different method is now used for sending those parameters, which doesn't require quoting and escaping.
18. Remove support for so-called "premature execution"
    
Premature execution was a method for getting information about a result set of a query, before SQLExecute() was called. That was always dangerous, as any side-effects of the query would happen, even if the application did not call SQLExecute() after all. With protocol version 3, there is a safer way to do this. This also removes the DisallowPremature configuration option, as the driver never does premature execution anymore.
19. Fix buffer overrun when the server reports an error message longer than 4096 bytes
    
If the server returned an error longer than 4096 bytes, we would overrun the buffer by two bytes. If you're unlucky, that could lead to a crash. That consistently happened on Windows.
20. Improve parsing of INSERT INTO statements, for @@identity support
    
The code that parses INSERT INTO statements, to extract the name of the target table, got confused by values containing dots, and other things. Improve the parsing support to handle a wider range of INSERT statements.
21. Fix crash in SQLTables() function, if search_path was set to a non-existent schema
    
The server-side current_schema() function returns a NULL if search_path is set to a non-existent schema, and the driver code was not prepared for a NULL result.
22. Fix buffer overrun when connecting, and KeepaliveTime and KeepaliveInterval were both set
    
If the server returned an error longer than 4096 bytes, we would overrun the buffer by two bytes. If you're unlucky, that could lead to a crash, and it consistently did on Windows. Per report from Andrus Moor.

23. Set an error message when password is required to connect.
    
In previous versions, if a connection fails because the server requires a password and no password was supplied. Per report from "PSequel Support".

24. Improve parsing of INSERT INTO statements to support @@identity
    
Patch by Christian Ullrich.

25. Fix SQL_DATA_AT_EXEC processing for large objects.
    
Reported by Vadym Krevs.

26. The server decide if SQL_LONGVARBINARY means a bytea or large object.
    
This reverts the behaviour of SQL_LONGVARBINARY parameters.

27. Fix handling of return value in {? = CALL ... } syntax.
    
When sending the query to the server, need to use the number of parameters calculated by build_libpq_bind_params(), which excludes the OUT parameter, rather than the original number of parameters. Reported by Haribabu Kommi.

psqlODBC 09.03.0400 Release

1. Rip out broken retry/timeout logic in SOCK_wait_for_ready.
   
At a quick glance, the logic looked like retry with progressive timeout, but it was horribly broken. First of all, the timeout was only used if retry_count was passed as 0. And in that case, the timeout was 0. So retry_count == 0 actually meant "check if data can be read/written without blocking", while retry_count > 0 meant "sleep until data is available". There was also some code to handle retry_count < 0, but that was dead code because none of the callers passed a negative value. The hardcoded max. retry count of 30 retries was quite bogus too. Filling the output buffer more 30 times is not necessary an error, and it's not clear that the ssl library could not return SSL_ERROR_WANT_READ or SSL_ERROR_WANT_WRITE more than 30 times. Unlikely, yes, but possible. So, just rip out the retry_count and progressive timeout logic. Replace the retry_count argument with a simple "should block?" argument. I didn't dare to remove the condition that we always wait if SSL is enabled, although I don't understand that at all. But perhaps there's some well-hidden reason for that. Per report from Prakash Itnal, although this just removes the broken code, and won't enable the kind of a timeout that he wanted.
2. Clean up Windows makefiles.
   
Get rid of the CPU and CPUTYPE variables in win64.mak. Instead, assume that the TARGET_CPU variable has been set correctly in the environment. TARGET_CPU is set by "setenv" script that comes with the Microsoft Visual Studio C compiler package. It never worked to pass a "CPU=x86" and using a x64 compiler, for example, so seems better to determine the correct target from the environment, than fail with a cryptic error message if the CPU argument and the environment don't match. To build 32-bit binaries with win64.mak, simply do "setenv /x86" before running "nmake -f win64.mak". Similarly, to build 64-bit binaries, use "setenv /x64". It should now be possible to build both 32-bit and 64-bit binaries with win64.mak. Ideally, we could now remove win32.mak altogether, but there are small differences between win64.mak and win32.mak in the build options that they use. I don't understand the differences enough to dare to remove win32.mak yet. The CPU=AMD64 alias for x64 is no longer supported (as the whole parameter is now gone). Also do whitespace & other cosmetic fixes to win32.mak and win64, to make them more similar. This helps when comparing the two with "diff". Remove remains of the long-gone "MULTIBYTE" setting (it was misspelled in win64.mak, anyway)
3. Fix crash if connection is closed while sending a query to the server.
   
Some functions reset conn->sock to NULL on a socket error, so we must not keep a copy of it in a local variable. Also, SOCK_get_id should check for NULL argument like all the rest of the SOCK_get/put_* functions.
4. Rewrite the conversion functions between strings and SQL_NUMERIC_STRUCT.
   
Aside from making the functions more readable and faster, this fixes a number of bugs:
1. When converting a large number from decimal to binary, the least significant byte might be wrong.
   reported by Walter Couto.
2. In binary->decimal conversion, a number with small precision but large scale caused a "Floating point exception".
   
3. In binary->decimal conversion, building the final string was wrong, picking up "digits" from outside the reserved stack space, if scale was larger than 40
   
4. The param_string buffer allocated for the result of binary->decimal conversion was too small.
   
Also add a test cases for all of these bugs.
5. Fix ODBC function escape for SPACE function.
   
6. Make sure that the psqlodbc ANSI driver is ODBC3.5.
   
7. Register 64bit XA DLL on installation
   
Changes by a patch from Craig Ringer
8. Fix potentially uninitialized local pointer access which causes.
   
fatal compilation errors on vs2012/2013 reported by Craig Ringer.
9. Fixed ctype_length() is missing the length for SQL_C_BIGINT and SQL_C_NUMERIC.
   
patch by John Smith.
10. Fix access to unallocated memory in IAsyncPG.
    
IAsyncPG was deleting its self, then returning an instance variable by value. This is probably going to be safe pretty much all the time, but it's still a memory error, and DrMemory (a Windows valgrind-alike)
11. SQL_NUMERIC_STRUCT was not available until ODBC 3.0.
    
Also add a prototype for ResolveNumericParam for readability.
12. Increase digits used to convert from REAL/DOUBLEs to strings so that the reverse conversions can recover original values.
    
Patch originally by John Smith.
13. Clear errors for prepared statement for subsequent exections.
    
14. Fix the problems using SQLFetch on prepared INSERT with RETURNING clause.
    
reported by John Smith.
15. Fix use of server-queried version before connection.
    
psqlODBC attempted to use the version that's queried from the server before it was obtained from the server during the connection setup. Because the server version is initialized to 0.0, tests for "server version >= 7.4" or "server version >= 6.4" using the PG_VERSION_LT macro were always returning false. Instead, use PROTOCOL_64 and PROTOCOL_74 tests against conninfo. These don't offer any "greater than" form, so just test both. There's little point adding anything prettier when it can all be removed when v1/v2 protocol support is removed soon anyway. The main symptom of this issue was that psqlODBC was sending an invalid protocol message for the v3 protocol, "Q \0", causing the server to wait for the second half of the length word forever. At connection close, the server would emit:
  unexpected EOF within message length word
This issue was observed while tracking down problems with XA transactions in MSDTC failing to commit after a successful prepare because an isolated tx created via getLockedXAConn() got stuck here, but is not confined to that problem.
16. Don't use PG_VERSION_xx() macros in original_CC_connect().
    
it's not obtained from the server yet. This may fix another cause of the problem Craig Ringer reported.
17. Fix bug in SQLCancel().
    
If one thread was busy executing a savepoint statement, while another thread calls SQLCancel() on the statement, SQLCancel() would try to release the locks held by the other thread. That's extremely racy; if the other thread tries to increment/decrement the lock_CC_for_rb counter at the same time, we could end up releasing the lock to many times or too few times. Also, releasing a pthread mutex from a different thread than the one who acquired it is undefined behavior. AFAICS the DiscardStatementSvp() call in SQLCancel() can be safely removed. It's outright wrong and dangerous if the statement is busy executing in a different thread, and the other codepaths in PGAPI_Cancel() already call DiscardStatementSvp(). Bug report by Jade Koskela.
18. Fixes a NULL pointer dereference occurring when connecting.
    
using SSPI authentication (without kerberos, a local Windows user trying to connect to postgres instance on same host). Patch originally by Nikhil R Deshpande.
19. Fix bug with UseDeclareFetch=1 when a transaction is committed before fetch.
    
If a server cursor is closed before the application has fetched any rows from the result set, the "base" of the result set's cached rowset was off-by-one. Also add a regression test for the same. This fixes the bug reported by Jan-Peter Seifert.
20. Fix driver name mismatch between 32-bit ODBC app and 64-bit MSDTC host.
    
If the 32-bit driver runs on a 64-bit host within a 32-bit app running under SysWow64 then XA transactions will fail to recover. That's because recovery is done by pgxalib.dll which runs with in MSDTC.exe - a 64-bit application. It connects to a DSN supplied by psqlODBC in EnlistInDtc_1pipe during DTC enlistment by the XA resource manager. This DSN uses the driver name of the client's ODBC driver - which is fine so long as the client and MSDTC architectures match. If they don't then MSDTC will fail to connect to Pg, so it'll never resolve transactions. Traces of MSDTC will show RM_COMMIT_DELIVERY_FAILED_DUE_TO_CONNECTION_DOWN events. Traces of pgxalib.dll will show xa_recover(..) being entered, followed by "SQLDriverConnect return=-1".
21. Remove PgDtc_set_property(.., prepareRequestded, 0) which causes a crash in case PREPARE TRANSACTION error.
    
The property prepareRequested isn't set anywhere.
22. Add KeepaliveTime and KeepaliveInterval options to specify time or interval for keepalive.
    
23. Fix a memory leak check qresult.
    
Per report by Pluto Cobain.
24. Get rid of copy and paste code in installer/ and unify them.
    
The powershell scripts had lots of duplication. Make the unified code instead. Original patch was provided by Craig Ringer.
25. Clear unused stuff from PgDtc_isolate() function.
    
Also add CC_initialize_pg_version() call after CC_copy_conninfo() so as to avoid the trouble caused by early PG_VERSION_xx() calls.
26. Fix race condition in getting the current decimal separator.
    
The code to get the current decimal separator was not thread-safe. If two threads concurrently saw that the locale had changed, they might both try to free() the same string. This isn't a perfect fix, there's still a race condition if a thread changes the locale while another thread is running the psqlodbc code. But at least this eliminates the race condition when the locale was changed earlier, not concurrently, and even if it is changed concurrently, this is less likely to crash, depending on the way localeconv() is implemented in libc.
27. Fix bug in building an abbreviated connection string.
    
Introduced by the keepalive-patch, which accidentally removed a line from sprintf format string. Caught by a compiler warning.
28. Fix buffer overflow in SQLGetData.
    
If a PostgreSQL timestamp column is fetched using SQLGetData, into a SQL_C_CHAR output variable, with size of 20-22 bytes, the driver might overrun the buffer by a few bytes with unusual timestamp values, with year < 0 or > 10000. While at it, fix the truncation behavior for date/time/timestamp values so that if the output buffer is too small, we still write what fits and truncate, instead of writing nothing. This is something that one might do on purpose, e.g. if you only want to fetch the year part of a date, you might call SQLGetData() on a date column, with only 4 bytes long output buffer. (although we haven't heard any complaints from the field). Also, if the output buffer is of type SQL_C_WCHAR, but the buffer length is odd, and we have to truncate, we didn't NULL-terminate the string. Using an odd-sized buffer with SQL_C_WCHAR is a pretty unlikely thing to do, but nevertheless I think this is more correct behavior.
29. The function SC_setInsertedTable() prepares an info for subsequent 'select @@IDENTITY' command.
    
Take table name like catalog.schema.table into account in the function. Fix the bug reported by Phillippe Champignon.
30. Improve pgxalib.dll(MSDTC support).
    
1. Add an option which lets xa_open() try to connect to the database immediately. This would cause an error at enlistment in dtc when the authentication method depends not only on the db user but also on the OS user, which means the automatic transaction recovery is unavailable.
   
2. Log the error message in case of SQLDriverConnect() error.
   
3. Improve the logging.
   
31. Fix the regkey search bug when XARMcreate() causes an error at enlistmentvin DTC.
    
32. Additonal improvement on MSDTC enlistment.
    
Unfortunately the current pgxalib.dll can't recover transactions which use sslmode verify-[ca|full] or whose authentication is SSPI, certificate or ident. When MSDTC.exe tries to connect to PostgreSQL it does so under Windows user NETWORKSERVICE, which won't match the username the original user connected to PostgreSQL as. So PostgreSQL will reject the connection.
33. Fix SQLTables column names to be ODBC version 3 compliant.
    
In ODBC3, TABLE_QUALIFIER was renamed to TABLE_CAT, and TABLE_OWNER to TABLE_SCHEM. Per report David Hedberg.
34. Add a new dialog page to the setup program and allow the setting of PREFERLIBPQ and XAOPT option.
    
35. Add a powershell script buildBootStrapper.ps1 which builds bootStrapper program with some optional parameters.
    
Quote the version number of bootstrapper from the configuration xml file unless specified. Also change the name of 32bit MSI in the bootstrapper according to changes of 32bit installer build system. Change buildInstallers.ps1 build the bootstrapper together. Change ALLUSERS from "2" to "1" of the 64bit installer.

psqlODBC 09.03.0300 Release

1. Fix implicit casts between SQLCHAR and char. Don't rely on -Wno-pointer-sign.
   
2. Pass the argument to isalpha/isspace etc. routines as unsigned char.
   
Per the C standard, the routine should be passed an int, with a value that's representable as an unsigned char or EOF. Passing a signed char is wrong, because a negative value is not representable as an unsigned char. Unfortunately no compiler warns about that.

3. Use "function(void)" instead of "function()" to declare 0-arg functions.
   
"function(void)" is the correct syntax in a function declaration, although in practice compilers accept the latter too. It's OK in a function definition, but change those too for consistency. This fixed by Michael Paquier.

4. Add missing function declarations.
   
These are not actually used outside of pgtypes.c, but let's be consistent and declare them in pgtypes.h This fixed by Michael Paquier.

5. Use the PG_CONFIG setting from ./configure cmd line in regression tests.
   
If compiling without libpq (configure --without-libpq), then you still need to have pg_config in path to build the regression tests, or specify PG_CONFIG at the make command line.

6. Add more test cases.
   
Add a new test case to test ODBC functions deprecated in ODBC 3.0, ODBC catalog functions, SQLGetConnectOption and Avoid deprecated SQLAllocStmt/SQLFreeStmt in regression tests. and more. Per Michael Paquier

7. Don't clear error number when SQLGetDiagRec is called.
   
Calls to SQLGetDiagRec are supposed to be nondestructive, per ODBC spec. Also add a test case for that.

8. Fix two UseDeclareFetch bugs.
   
1. NOTICE messages were not delivered to the application, if they arrived as response to the DECLARE CURSOR statement.
2. Array-bound parameters on SELECT-queries caused a "cursor already open" error.
These bugs were found by running the regression suite with UseDeclareFetch=1. It's now clean.

9. Fix crash if connection is closed during CC_send_query_append function.
   
This fixed Malcolm MacLeod

psqlODBC 09.03.0210 Release

1. SSL verify[-(ca|full)] is avaiable since 8.4.
   
There seems no need to check it. Also there's no need to call lt_dlopen currentl y.

2. Update EXTRA_DIST in Makefile.am
   
Files for the new lfconversion test case were missing.

3. Fix locking in SC_set_prepared
   
added ENTER/LEAVE_CRIT_CS calls in SC_set_prepared.

4. Add test case for CTE queries (WITH ...)
   
None of the existing test cases covered that. This test case gives a different e rror message than the original one, but it'sthe same underlying issue.

5. Silence compiler warning.
   
SQLGUID format and conn_settings cannot be NULL.
This fixed by Michael Paquier.

6. Fixed POSTGRES_RESOURCE_VERSION variable.
   
Pass the content of POSTGRES_RESOURCE_VERSION variable to the resource compiler correctly so that FileVersion and ProductVersion are properly set.

7. Change regression test positioned-update
   
Use "exit(1);" instead of "return;" to exit from main().
This fixed by Michael Paquier, at least some compiler on OS X didn't like it.

psqlODBC 09.03.0200