Re: Infinite loop on error

On Mon, Aug 2, 2010 at 7:56 AM, Marko Kreen <[email protected]> wrote:
> On 8/2/10, Daniele Varrazzo <[email protected]> wrote:
>> On Wed, Jul 28, 2010 at 4:54 PM, Marko Kreen <[email protected]> wrote:
>> =A0> Psycopg has several loops in form:
>> [...]
>> =A0> The problem is that if libpq gets so fatal error it decides
>> =A0> to close the connection, PQgetResult() will never return NULL
>> =A0> from that connection, thus infinite loop.
>>
>> Are you sure about that? I don't have the libpq source code handy, I
>> =A0will surely check that.
>
> Yeah, found out via strace/gdb... =A0strace was showing infinite
> close(-1) - thus libpq was trying to close already closed connection.
>
>> But, if it doesn't return a NULL with the
>> =A0connection in broken state... what does it return?
>
> Probably PQmakeEmptyPGresult(conn, PGRES_FATAL_ERROR)

Yup, seen in the libpq source.


>> =A0PQgetResult() also returns null when there is no result to fetch. I
>> =A0may have extrapolated excessively assuming that it returns null when
>> =A0the connection is closed too, but it doesn't seem an excessively
>> =A0stretched interpretation.
>
> Yeah, seems NULL means "query is finished _and_ everything is fine".

I shall propose a clarification to the pg docs ml.


>> =A0I'm curious to do some test in the next days.
>
> Have fun.

I did, but I've not been able to reproduce a loop (I've produced
segfaults, crashes... but not an infinite loop). Do you have a recipe
to reproduce it? I have a patch adding the proposed check to the loops
(http://tinyurl.com/3463bql). If you can test it (or propose how to
test it) it would be great.

Thank you, cheers.

-- Daniele