Hi Dave,

I think it should solve this problem. Thank you for your help.

It seems that there won't be a problem with use of unsigned values, since all values that could be negative are handled.
I'm not sure if "stapos" should be short.   Product of 2byte*2byte variables could require 4 bytes to hold the value, 
if it's possible for "error->recsize" and "RecNumber" to hold big enough values at the same time.
Maybe make "stapos" signed/unsigned int to make sure it won't happen?

Regards,
Andrey

27.02.2024 21:21, Dave Cramer wrote:

Hi Andrey,

Thoughts use unsigned word for lengths to avoid overflow by davecramer · Pull Request #3 · postgresql-interfaces/psqlodbc (github.com)

Dave Cramer

www.postgres.rocks

On Tue, 27 Feb 2024 at 01:07, Andrey Sukhanov <siwenter@gmail.com> wrote:

Hi Dave,

Thank you for your time.
The problem is caused by signed integer overflow, which leads to attempt to read from unallocated memory.

I have downloaded psqlodbc 16 and debug symbols from https://www.postgresql.org/ftp/odbc/versions/.
Source code from for psqlodbc 16 from https://git.postgresql.org/git/psqlodbc.git.
I compiled the attached program with msvc 14 and msvc 16 compilers (same results). Program should be compiled with defined "UNICODE" to use functions (SQLGetDiagRecW etc.) that accept wchar.

On windows 10 attached program crashes inside ER_ReturnError line environ.c:249. Where it tries to memcpy from the buffer "msg" with offset "stapos".


"msg" is a valid pointer to "pgerror->__error_msg". "pgerror" also have field "pgerror->errorsize" with value "32690".
Offset "stapos" is calculated as "stapos = (RecNumber - 1) * error->recsize;" 
Here "error" is "pgerror", "recsize" = 99, "RecNumber" = 332.
"stapos" should be assigned the value of 32769  but it has type SWORD (short int) (INT16_MAX is 32767).
Expression"if (stapos > msglen)" (here "msglen" = 32690) evaluates to "false", because of signed integer overflow.  It should've been "true" instead, and function would've returned SQL_NO_DATA_FOUND.
Callstack:
     vcruntime140.dll!memcpy() line 438
     psqlodbc35w.dll!ER_ReturnError(PG_ErrorInfo * pgerror, short RecNumber, unsigned char * szSqlState, long * pfNativeError, unsigned char * szErrorMsg, short cbErrorMsgMax, short * pcbErrorMsg, unsigned short) line 250
     psqlodbc35w.dll!PGAPI_StmtError(void * hstmt, short RecNumber, unsigned char * szSqlState, long * pfNativeError, unsigned char * szErrorMsg, short cbErrorMsgMax, short * pcbErrorMsg, unsigned short) line 1626
     psqlodbc35w.dll!PGAPI_GetDiagRec(short) line 57
     psqlodbc35w.dll!SQLGetDiagRecW(short fHandleType, void * handle, short iRecord, wchar_t * szSqlState, long * pfNativeError, wchar_t * szErrorMsg, short cbErrorMsgMax, short * pcbErrorMsg) line 234
     odbc32.dll!SQLGetDiagRecW()
     odbc.pg.exe!CollectDiagRecords(short hndlType, void * hndl) line 30
     odbc.pg.exe!main(int argc, char * * argv) line 117

That means, that the problem is caused by signed integer overflow, which in turn causes memcpy to read from unallocated memory.
The easiest way to reproduce this problem is to try increasing the number of "raise notice" messages in for loop. If their number gets past certain point, no diagnostic records from SQLGetDiagRec are being returned at all. Then I've decreased the number of these messages until they started appearing again. In my case at this number I am able to reproduce the problem. With further decrease of this number problem doesn't appear anymore.

I can't fix this problem myself, because I'm not familiar with the project, and I don't know how to properly test psqlodbc on different platforms.

Regards,
Andrey

27.02.2024 3:11, Dave Cramer wrote:

Hi Andrey,

I've tried on PostgreSQL 11 and 15 with no luck to replicate this problem.

Is it possible for you to debug it and send more information?

Dave Cramer

www.postgres.rocks

On Thu, 22 Feb 2024 at 01:29, Andrey Sukhanov <siwenter@gmail.com> wrote:

Dear pgsql-odbc developers,

Windows 10,  psqlodbc 16 (psqlodbc35w.dll), postgresql 11.
Getting certain amount of diagnostic records with SQLGetDiagRecW crashes
the application with memory access violation.

Steps to reproduce:
1. Create procedure:
CREATE OR REPLACE PROCEDURE crashme()
LANGUAGE plpgsql
AS $$
BEGIN
FOR i IN 1..841 LOOP
         RAISE NOTICE 'msgmsgmsgmsg (%)', i;
END LOOP;
END; $$;

2. Use example code in attachments.
3. Application crashes with memory access violation after calling
SQLGetDiagRecW function, with iRecord = 332.
Application doesn't crash if number of iterations in  procedure's for
loop is changed.
Expected outcome: SQLGetDiagRecW would return SQL_NO_DATA when there's
no more diagnostic records.

Regards,
Andrey