Обсуждение: Buffer overrun in handle_notice_message()

Hi all,

I'd like to report a buffer overrun in handle_notice_message().
Analysis: when I run a query >4096 characters that has a 'C' at a
specific location (at a specific, small offset before a multiple of
4096), then the stack is trashed and the driver crashes. The code fragment:

for (;;)
{
    truncated = SOCK_get_string(sock, msgbuffer, sizeof(msgbuffer));
    if (!msgbuffer[0])
        break;

    mylog("%s: 'N' - %s\n", comment, msgbuffer);
    qlog("NOTICE from backend during %s: '%s'\n", comment, msgbuffer);
    switch (msgbuffer[0])
    {
        case 'S':
            strncat(msgbuf, msgbuffer + 1, buflen);
            strncat(msgbuf, ": ", buflen);
            buflen -= (strlen(msgbuffer) + 1);
            break;
        case 'M':
            strncat(msgbuf, msgbuffer + 1, buflen);
            msg_truncated = truncated;
            break;
        case 'C':
            if (sqlstate && !sqlstate[0] && strcmp(msgbuffer + 1, "00000"))
                strcpy(sqlstate, msgbuffer + 1);
            break;
    }
}



The problem is that PostgreSQL replies to the query with a "LOG"-type
message, quoting the _complete_ query. This loop retrieves the message
in chunks of 4096 bytes (i.e., sizeof(msgbuffer)) and normally just
outputs things to mylog and qlog, _except_ if the first character of a
chunk just happens to be 'S', 'M' or 'C'. In the 'S' or 'M' cases,
nothing bad happens, but in the 'C' case, the
strcpy(sqlstate,msgbuffer+1) copies 4 kbytes of data into the 8-byte
sqlstate buffer.

The way I understand it, the switch on msgbuffer[0] should only happen
on the first iteration, not on subsequent iterations. After I changed
the loop like that, my application ran without any problems.

Oh, in case you're wondering: protocol is 7.4+, PostgreSQL server
version is 8.1.4.

Cheers,
Bart

Bart Samwel wrote:
> Hi all,
>
> I'd like to report a buffer overrun in handle_notice_message().
> Analysis: when I run a query >4096 characters that has a 'C' at a
> specific location (at a specific, small offset before a multiple of
> 4096), then the stack is trashed and the driver crashes. The code
> fragment:
>
> for (;;)
> {
>     truncated = SOCK_get_string(sock, msgbuffer, sizeof(msgbuffer));
>     if (!msgbuffer[0])
>         break;
>
>     mylog("%s: 'N' - %s\n", comment, msgbuffer);
>     qlog("NOTICE from backend during %s: '%s'\n", comment, msgbuffer);
>     switch (msgbuffer[0])
>     {
>         case 'S':
>             strncat(msgbuf, msgbuffer + 1, buflen);
>             strncat(msgbuf, ": ", buflen);
>             buflen -= (strlen(msgbuffer) + 1);
>             break;
>         case 'M':
>             strncat(msgbuf, msgbuffer + 1, buflen);
>             msg_truncated = truncated;
>             break;
>         case 'C':
>             if (sqlstate && !sqlstate[0] && strcmp(msgbuffer + 1,
> "00000"))
>                 strcpy(sqlstate, msgbuffer + 1);
>             break;
>     }
> }

Hi Bart,
Hmm, ISTM I should place the following code at the end of the above loop
not after the above code..

               while (truncated)
                        truncated = SOCK_get_string(sock, msgbuffer,
sizeof(msgbuffer));

Actually I do so In handle_error_message().
I would fix it ASAP.
Thanks.

regards,
Hiroshi Inoue