Date: Tue, 22 Feb 2022 21:00:15 -0800
From: Noah Misch <noah@leadboat.com>
To: "egashira.yusuke@fujitsu.com" <egashira.yusuke@fujitsu.com>
Cc: "'pgsql-bugs@lists.postgresql.org'" <pgsql-bugs@lists.postgresql.org>
Subject: Re: Reconnect a single connection used by multiple threads in
 embedded SQL in C application causes error.
Message-ID: <20220223050015.GC3838703@rfd.leadboat.com>
References: 
 <TYWPR01MB7202E14319002430B83F7ABBFF3A9@TYWPR01MB7202.jpnprd01.prod.outlook.com>
 <20220221212928.GB3799825@rfd.leadboat.com>
 <TYWPR01MB7202664D55ED1F8103B42FA7FF3B9@TYWPR01MB7202.jpnprd01.prod.outlook.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: 
 <TYWPR01MB7202664D55ED1F8103B42FA7FF3B9@TYWPR01MB7202.jpnprd01.prod.outlook.com>
User-Agent: Mutt/1.5.24 (2015-08-30)
Archived-At: 
 <https://www.postgresql.org/message-id/20220223050015.GC3838703%40rfd.leadboat.com>
Precedence: bulk

On Tue, Feb 22, 2022 at 12:06:23PM +0000, egashira.yusuke@fujitsu.com wrote:
> > > And, is there limitation that we can't CONNECT or DISCONNECT the DEFAULT connection inside a thread in multithreaded application?
> > 
> > For now, yes.  This is a bug.  The documentation you quoted is out of date
> > with respect to the code.  That documentation comes from commit 8f61184
> > (2003-12).  The behavior of connections shared across threads changed in
> > commit 757fb0e (2004-03).  Since that change, there's a per-thread record of
> > the default connection, in addition to a global notion of the default
> > connection.  I suspect a few use cases work well today:
> > 
> > - One thread does all the CONNECT and DISCONNECT commands.  Other threads just
> >   use the default connection, with mutual exclusion.
> > - Each thread does its own CONNECT, DISCONNECT, and other commands.  Each
> >   thread has its own default connection.  No sharing at all.
> 
> The documentation from commit 8f61184 seems to have been updated by the commit 0f33ee0 (2017-06).
> However, there is still lacking of information about the DEFAULT connection on per-thread, I understood.

Yes.

> > Your example does the CONNECT to establish the default connection in one
> > thread, and it does the DISCONNECT of the default connection in another
> > thread.  That's buggy today.
> > 
> > Another behavior that may qualify as a bug: ECPGsetconn() updates only the
> > thread-specific connection, while ECPGconnect() updates both the
> > thread-specific and global connections.
> 
> I also understood that my example is impossible by the implementation bugs.
> 
> I think it would take some time to fix these bugs. 

Yes.  I suspect the hard part will be picking a specification that makes your
case work better without breaking other important cases that work today.

> If it takes long time to fix it, will this limitation(*) be added to the documentation?
> 
> [*] Based on your explanation:
>     "If we use the default connection inside a thread in multithreaded application, 
>      CONNECT and DISCONNECT commands are allowed only following cases.
>      - One thread does all the CONNECT and DISCONNECT commands. 
>        Other threads just use the default connection, with mutual exclusion.
>      - Each thread does its own CONNECT, DISCONNECT, and other commands. 
>        Each thread has its own default connection. No sharing at all."

Would you like to propose a patch?

> An additional report: 
> I also ran my reproduction application built with rh-postgresql12 libraries (installed from Red Hat Software Collections), 
> and it crashed instead of an error in step 5 in my first email.
> 
> * Actually, the error message in my first email came from postgresql built by myself (./configure --prefix=xxx).
> 
> The backtrace is following.
> (gdb) bt
> #0  0x00007fa11a4ddd0f in ecpg_find_prepared_statement () from /opt/rh/rh-postgresql12/root/usr//lib64/libecpg.so.rh-postgresql12-6
> #1  0x00007fa11a4de11b in ecpg_prepared () from /opt/rh/rh-postgresql12/root/usr//lib64/libecpg.so.rh-postgresql12-6
> #2  0x00007fa11a4d8b65 in ecpg_do_prologue () from /opt/rh/rh-postgresql12/root/usr//lib64/libecpg.so.rh-postgresql12-6
> #3  0x00007fa11a4d8c33 in ecpg_do () from /opt/rh/rh-postgresql12/root/usr//lib64/libecpg.so.rh-postgresql12-6
> #4  0x00007fa11a4d8d5b in ECPGdo () from /opt/rh/rh-postgresql12/root/usr//lib64/libecpg.so.rh-postgresql12-6
> #5  0x0000000000401143 in thread_main2 ()
> #6  0x00007fa11a2bdea5 in start_thread () from /lib64/libpthread.so.0
> #7  0x00007fa119fe68cd in clone () from /lib64/libc.so.6
> 
> I think this happened because the PGconn referenced by Thread#1 was destroyed in step 4 and then overwritten with invalid data.

That makes sense.