NULL data not mapped to NA with RODBC on 64-bit Mac OS X

Hi Harlan,

I don't think this is a right way to fix.
Because
(int64_t) 0xFFFFFFFF == 4294967295 != -1
the bug resides on the code setting
ColData[i].IndPtr[row].
Changing the way to interpret it is a wrong solution.
Though, it might work as a workaround for the bug in other parts.

The commented-out warning() in the code was used to confirm that without
the casts,


You did see only the lower 32 bits?

--

Tomoaki NISHIYAMA


Advanced Science Research Center,

Kanazawa University,

13-1 Takara-machi,

Kanazawa, 920-0934, Japan


On 2011/04/19, at 1:31, Harlan Harris wrote:

OK, with the help of a rusty C coder at work, I've confirmed that the
solution the Python crew figured out also seems to work for me. I have no
idea if it breaks the 32-bit MacOS build, or any other build. Note that this
seems to be a very subtle effect of the way the compiler and header files
are interacting, as far as I can tell. The commented-out warning() in the
code was used to confirm that without the casts, the len variable and
SQL_NULL_DATA are, in fact, the same, and that despite their being the same,
the equality test fails.

A patch is attached. All of the len variables are cast to ints. (Note that
casting to long, which should work, doesn't.) Would someone please apply to
patch to the source tree and do some regression testing?

-Harlan


On Thu, Mar 31, 2011 at 9:35 AM, Harlan Harris <harlan at harris.name> wrote:

Prof. Ripley? Anyone? Should I cross-post this to R-Help? If someone would
be willing to fix this for pay, please email me a proposal and I'll see if
my employer would be willing to sponsor a fix.

 -Harlan


On Tue, Mar 29, 2011 at 9:43 AM, Harlan Harris <harlan at harris.name>wrote:

Following up on this... Prof. Ripley? Is this something you can address?

 -Harlan


On Thu, Mar 24, 2011 at 9:54 AM, Harlan Harris <harlan at harris.name>wrote:

Thanks for the reply, Marc!

Ah, it does appear I'm booting in 32-bit mode. From the terminal:

uname -a

Darwin Harlan-Harris-MacBook-Pro.local 10.6.0 Darwin Kernel Version
10.6.0: Wed Nov 10 18:13:17 PST 2010; root:xnu-1504.9.26~3/RELEASE_I386 i386

I can usually run stuff in 32-bit mode, but sometimes I would like the
ability to use the extra RAM. Also, certain interfaces to R (such as the new
RStudio) will run R in 64-bit mode, so for now, I can't use RStudio when I'm
pulling data via ODBC that might have NULLs. I think Rscript also wants to
run in 64-bit mode.

 -Harlan


On Thu, Mar 24, 2011 at 9:09 AM, Marc Schwartz <marc_schwartz at me.com>wrote:

Hi Harlan,

I suspect that we will need to wait for Prof. Ripley to be able to look
into this further. There may be subtle issues at play here and I would defer
to his more intimate knowledge of the ODBC tool chain.

I do know that the Actual driver is both 32 and 64 bit in a single
installation. However, I run R in 32 bit mode (don't need the extra address
space, even though I now have 8Gb on my MacBook Pro), so have not
encountered the 64 bit issues that you report here.

I have the same header file. Note that SQLLEN is a signed declaration
below. It is defined as a 'long' generally and as a INT64 for 64 bit
Windows. It would be 'unsigned long' or UINT64, respectively for an unsigned
value.

One question for you however, which is do you know if you are booting
the 32 bit or 64 bit OSX kernel on your Mac? Macs prior to mid-2010 (like
mine) running Snow Leopard typically default to the 32 bit kernel, though
can be 64 bit 'capable'. It depends upon various factors including the
presence of a 64 bit EFI. Newer Macs now boot into the 64 bit kernel by
default. I have seen some problems reported on older Macs where the default
kernel was changed from 32 bit to 64 bit, resulting in subtle and not so
subtle problems with some drivers. Booting to the 32 bit kernel still allows
you to run 64 bit apps under Snow Leopard, just that the kernel and
extensions are 32 bit.

Regards,

Marc


On Mar 23, 2011, at 2:29 PM, Harlan Harris wrote:

Any thoughts? Is there someone else or somewhere else I might ask?

-Harlan

On Tue, Mar 22, 2011 at 1:31 PM, Harlan Harris <harlan at harris.name>

wrote:

Hello,

I seem to have hit a bug in RODBC on 64-bit versions of R on Mac OS

X. I'm

running R 2.12.1, and using R to talk to Oracle and other databases

using

RODBC 1.3-2. In 32-bit mode (R32), everything works. (The ODBC

drivers are

the ones from Actual.) In 64-bit mode (R), it seems as if NULL data

is not

being treated as missing. For example, a NULL value in an integer

column is

being returned as a 0 instead, which is wrong/bad.

Diving into the code and Google a bit, it seems like an issue that's

been

seen before in other contexts. Here's the relevant C code, from
RODBCFetchRows:

       for(row = thisHandle->rowsUsed;
       row < thisHandle->rowsFetched && j <= maximum;
       j++, row++)
       {
       thisHandle->rowsUsed++;
       if(j > blksize) {
           blksize *= 2;
           for (i = 0; i < nc; i++)
           SET_VECTOR_ELT(data, i,
                      lengthgets(VECTOR_ELT(data, i), blksize));
       }
       for (i = 0; i < nc; i++) {
           SQLLEN len = thisHandle->ColData[i].IndPtr[row];
           switch(thisHandle->ColData[i].DataType) {
           case SQL_DOUBLE:
           REAL(VECTOR_ELT(data, i))[j-1] =
               len == SQL_NULL_DATA ? NA_REAL :
               thisHandle->ColData[i].RData[row];
           break;

etc...

Googling for SQL_NULL_DATA and related terms finds this page, about

the

same issue in a Python driver:

http://code.google.com/p/pyodbc/issues/detail?id=51

It appears as if SQLLEN might be unsigned in 64-bit builds, while
SQL_NULL_DATA is -1. Casting len to be an int rather than an

unsigned int

seems to solve the problem for them.

On my machine, /usr/include/sqltypes.h starts like this:

/*
*  sqltypes.h
*
*  $Id: sqltypes.h,v 1.23 2007/10/07 13:27:13 source Exp $
*
*  ODBC typedefs
*
*  The iODBC driver manager.

and defines SQLLEN as follows:

#ifdef _WIN64
typedef INT64                   SQLLEN;
typedef UINT64                  SQLULEN;
typedef UINT64                  SQLSETPOSIROW;
#elif defined(STRICT_ODBC_TYPES)
typedef long                    SQLLEN;
typedef unsigned long           SQLULEN;
typedef unsigned short          SQLSETPOSIROW;
#else
#define SQLLEN                  long
#define SQLULEN                 unsigned long
#define SQLSETPOSIROW           unsigned short
#endif

Seems to me like SQLLEN should be signed, but apparently not?

Can anyone help? Thank you!

-Harlan

<RODBC.c.diff>_______________________________________________

R-sig-DB mailing list -- R Special Interest Group
R-sig-DB at r-project.org