Re: Problem with procedures returning a SYS_REFCURSOR which is not open/executed - possible fixes

On 11/01/2013 21:22, John Scoles wrote:
> Hmm you sure pick the head scratchers.
>   
> My first thoughts where why whould you make such a procedure?? But then I relaisze it could have some use when prototying/develpopeing ect.
>   
No, this is production.

Sometimes there is a result-set to return and sometimes there is not. 
Our situation here is perhaps a little unusual in that there is no read 
or write access to any table in the database for anyone other than the 
owner of the package - definer rights. As a result there is a procedure 
to return any result-set the outside code needs. In this case the 
outside knows some work needs to be done but not what it is and usually 
does this:

alerted some work needs to be done
calls proc to get the work
depending on the work it might call another proc
once work is done calls a proc to say it is complete and audit it

I'm skipping the last 2 steps for some work by doing the dispatching in 
the database thus cutting out to 2 trips to the database. The downside 
is some of the work returns a cursor and some does not.

> Looking at
>   
> 'DBD::Oracle attempts to describe statements which are initialized but not executed'
>   
> We might be able to do something here.  I remember working on a patch some time ago (ie 2008) to get the describe after the execute then do all the binding at the end of an execute etc as the newer versions of Oracle return this data for you.  Not sure it that is the right time for this on as if I remember correctly there was alot of reprogrmming involved.
>   
Doing the describe before the execute is a pain and the current describe 
code makes all the decisions on how columns will be bound which means a 
type to bind_col is ignored. Hence rt's like:

https://rt.cpan.org/Ticket/Display.html?id=71810
ora_type => ORA_RAW doesn't work for select

where the reporter has a raw column which is bound as a string so Oracle 
converts it to hex and the result is twice as long. Of course we could 
just change the default binding of raw to bytes but I was reticent to do 
that since the code had been that way a long time and people might be 
relying on it. If the describe was afterwards the reporter could just 
specify raw in the bind_col call thus not upsetting anyone else.

I got away with implementing strictly typed etc on bind_col because they 
are applied afterwards.

> Will have to look into that one.
>   
> Anyway I agree that pp_exec_rset is the place to make your change  as the is safe area to do it.  Perhaps we can delay the pre_exec part untill after the inital query is executed by that time you will know if you have a ref that you will need to bind and return??
That was one of my thoughts also. In actual fact, I'd rather have undef 
back for the cursor when there is no open cursor returned, it seems more 
intuitive to me and avoids creating a new sth for no purpose. My only 
concern was if it was at all possible to return an initialised cursor 
for a valid result-set which was just not executed yet - not that I know 
if you can do this or how. I will hopefully try and find the time to 
look at that.
>   
>   
> Been a while since I looked at this part of the code so bear with me if I sound a little rusty.
>   
> Cheere
> John
Anyway, the change I arrived at does not seem to have any negative 
affects on the test suite or our application so far. If I can find a 
better way of doing it which is not a total overhaul of dbd_describe I 
will but as I seem to be the only one with this issue I did not want to 
risk breaking other things.

Martin
>   
>   
>   
>
>   
>
>> Date: Fri, 11 Jan 2013 16:04:13 +0000
>> From: [email protected]
>> To: [email protected]
>> Subject: Problem with procedures returning a SYS_REFCURSOR which is not open/executed - possible fixes
>>
>> I am using DBD::Oracle and calling a procedure which returns a reference cursor. However, sometimes the reference cursor is not opened and only the procedure knows this. The problem is if I call the procedure from DBD::Oracle and the cursor is not opened I get an Oracle error saying the cursor is not executed:
>>
>> test procedure:
>> procedure p_n2(pcur OUT SYS_REFCURSOR) AS
>> begin
>> pcur := NULL;
>> end;
>>
>> example perl:
>> my $s = $h->prepare(q/begin mypkg.p_n2(?); end;/);
>> $s->bind_param_inout(1, \my $cursor, 100, {ora_type => ORA_RSET});
>> $s->execute; # errors
>>
>> The error occurs because DBD::Oracle attempts to call dbd_describe on the returned cursor (before perl land even sees it) and that code does things like call OCIAttrGet for PARAM_COUNT etc which Oracle disallows if the statement is not executed.
>>
>> An easy solution is to just open an empty cursor if the procedure cannot open a real one by doing something like:
>>
>> open pcur for select 1 from dual;
>>
>> but I don't like that as DBD::Oracle will make dozens of calls and do quite a bit of work in dbd_describe which is wasting time and the purpose of the change to my procedure is to speed this application up not slow it down.
>>
>> Just to be clear in case anyone thinks I've just invented a scenario and there is nothing actually wrong with DBD::Oracle - it is most definitely a bug for DBD::Oracle to attempt to describe a non-executed stmt.
>>
>> Possible solutions are complicated by these facts (some of which I only discovered after a few attempts of patching this):
>>
>> o DBD::Oracle magics a DBI sth into existence but it does this before the Oracle's execute is called in dbd_execute.
>>
>> o DBD::Oracle attempts to describe statements which are initialized but not executed.
>>
>> o DBD::Oracle will always call dbd_describe on any statement if it is not marked active when you attempt to access any attribute of the sth.
>>
>> So my first hack was to change dbd_describe to ask for the statement state and if it was initialized but not executed it just returns having done nothing else. This works because the sth is left in an Active state and dbd_describe has been called so the internal flag dbd_describe_done is set. However the down sides are a) you need to check to see if NUM_OF_FIELDS returns something before using it b) the sth is not really Active c) dbd_describe is a critical function and is used for all stmt handles, not just ones magicked into existence for returned cursors.
>>
>> My second attempt was to hijack the code in pp_exec_rset which is called just before execute to magic the stmt handle into existence and after execute to set the sth up and call dbd_describe. My thinking was that it was much safer changing code here. The first sub attempt to simply to check the stmt state and if initialized and not executed, don't call dbd_describe and don't set the sth Active. The idea was to check Active outside in perl land. It does not work because any time you attempt to access an attribute of a non-Active sth where dbd_describe has not been called, you guessed it, DBD::Oracle calls dbd_describe - so I am back where I started. My second sub attempt was to outright lie and set dbd_describe_done and leave Active off so from perl land I just need to test Active flag. This works and is a safer change since it ONLY applies to sth handles magicked into existence for returned cursors. Also, if you attempt to do anything else with the sth it errors as it should:
>>
>> DBD::Oracle::st fetch failed: ERROR no statement executing (perhaps you need to call execute first) at bz1245.pl line 16.
>>
>> Wondered if anyone else had any thoughts on this. Ideally I'd like a solution people are happy enough to go into DBD::Oracle officially as I maintained quite a number of bespoke patches here for some years in the past and it was a PITA. Also, the more simple the solution the better as the internals of DBD::Oracle are quite complex and I'd rather not re-engineer a load of code just for this.
>>
>> Martin
>> -- 
>> Martin J. Evans
>> Easysoft Limited
>> http://www.easysoft.com
>