Re: rollback to savepoint leads to transaction already in progress

On Thu, Oct 14, 2010 at 6:53 AM, Michael Meskes<meskes(at)postgresql(dot)org> wrote:

> You seem to be answering to an email that I didn't see and couldn't find in
> the archive either. Was it send in private or to the list? If it went to the
> list I might be lagging behind.
>

On 14/10/10 22:51, Robert Haas wrote:
> I didn't get it either. The original report was posted to pgsql-bugs
> on October 10.
>

It might be lag; it'll probably turn up, about 6 seconds after I press *S

-------- Original Message --------
Subject: Re: [BUGS] rollback to savepoint leads to transaction already
in progress
Date: Thu, 14 Oct 2010 10:44:50 +0900
From: Itagaki Takahiro<itagaki(dot)takahiro(at)gmail(dot)com>
To: David Newall<postgresql(at)davidnewall(dot)com>, PostgreSQL Hackers
<pgsql-hackers(at)postgresql(dot)org>
CC: pgsql-bugs(at)postgresql(dot)org

On Mon, Oct 11, 2010 at 7:42 AM, David Newall
<postgresql(at)davidnewall(dot)com> wrote:
> Trivial program to demonstrate problem:
>
> main() {
> ECPGdebug(1,stderr);
> exec sql connect to postgres;
> exec sql set autocommit to off;
> exec sql start transaction;
> exec sql savepoint s;
> exec sql rollback to s;
> exec sql commit;
> return 0;
> }
>
> Output:
> [28397]: ECPGdebug: set to 1
> [28397]: ECPGconnect: opening database postgres on<DEFAULT> port<DEFAULT>
> [28397]: ECPGsetcommit on line 4: action "off"; connection "postgres"
> [28397]: ECPGtrans on line 5: action "start transaction"; connection "postgres"
> [28397]: ECPGtrans on line 6: action "savepoint s"; connection "postgres"
> [28397]: ECPGtrans on line 7: action "rollback to s"; connection "postgres"
> [28397]: ECPGtrans on line 8: action "commit"; connection "postgres"
> [28397]: ECPGnoticeReceiver: there is already a transaction in progress
> [28397]: raising sqlcode -603
>
> Problem:
> It shouldn't raise "there is already a transaction in progress" error,
> particularly when doing a commit. Remove "rollback to s" and no problem.
>
> Environment:
> ecpg (PostgreSQL 8.4.5) 4.5.0
> This appears to be a regression; it doesn't occur with ecpg (PostgreSQL
> 8.3.8) 4.4.1

The bug comes from string-based transaction control in ECPGtrans().
The code cannot distinguish ROLLBACK TRANSACTION and ROLLBACK TO savepoint.
----
if (strncmp(transaction, "commit", 6) == 0 || strncmp(transaction,
"rollback", 8) == 0)
con->committed = true;
else
con->committed = false;
----

I think the string-comparison is unreliable. So, I'd like to replace
the code to use PQtransactionStatus(). I have two patches to do it:
The first one (ecpg-trans-quick_20101014.patch) is a quick fix
that replaces only the above test.
The second one (ecpg-trans-full_20101014.patch) replaces all of
struct connection->committed with PQtransactionStatus().

Which solution is better? Or, another idea?

--
Itagaki Takahiro