[mico-devel] Patch for GIOPConn(Reader) Thread local storage issue, comments?

[Karel Gardas]

Hello,

On Thu, 17 Aug 2006, Arne Pajunen wrote:

> Hello,
>
> Attached is a patch of a fix I made to an issue we encountered with the 
> handling of the reader TLS (Thread Local Storage) keys in GIOPConn.
>
> The issue was such that we had a call being made to a server, which returned 
> a reference to a generic resource. When the client receives this, Mico does a 
> _is_a query to verify it in the thread where the reply was received. In our 
> case this _is_a query went to another process with a different connection 
> which was just started by the server.
>
> The problem was that the newly started connection created it's reader 
> threads, but then the query in the old connections reader thread calls 
> ORB:wait() to wait for the reply, and it determined that this is a reader 
> thread so it should directly call the dispatcher to wait for the reply. 
> Unfortunately, in this scenario it was the wrong reader thread, and the new 
> connections reader thread was also happily reading. This ended up with a race 
> in do_read(), where two threads would enter with the do_read simultaneously 
> causing a unpredictable condition.
>
> Unfortunately I'm unable to produce a reasonable test case for it as this 
> occurred in a complex system, and was kind of rare and timing keen (sometimes 
> had to run the program in while loops until it occurred).
>
> The exact issue is that the static MICO::GIOPConn::is_this_reader_thread() 
> method verifies whether current thread is a reader, but doesn't care for 
> which connection. I fixed this by making it use the pointer to the GIOPConn 
> instance as the thread local keys value, so it can verify that it's a reader 
> for the current connection.
>
> I'm not sure if this patch is perfectly correct (or more like I'm not sure I 
> fixed the right problem.) and whether it has any side effects. I would be 
> interested in comments and feedback on this patch.

your information are very interesting. Could you also be so kind and tell 
us what concurrency model do you use on client side? I would also be 
interested to know more details about how is it possible you call _is_a 
operation from reader thread directly. IMHO this shouldn't happen at 
all...

Thanks,
Karel
--
Karel Gardas                  kgardas at objectsecurity.com
ObjectSecurity Ltd.           http://www.objectsecurity.com