db-derby-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Bryan Pendleton <bpendleton.de...@gmail.com>
Subject Re: Could someone give me some guidance on DERBY-5560
Date Sat, 31 Dec 2011 16:26:03 GMT
> Basically what is happening is that the LogicalConnection.close() is being called which
attempts to recycle the physical
> connection by calling ClientPooledConnection.recycleConnection(). At the same time ClientPooledConnection.close()
is being
> called which attempts to call LogicalConnection.nullPhysicalConnection(). The first thread
holds a lock on LogicalConnection and
> needs the lock on ClientPooledConnection and the second thread holds a lock on ClientPolledConnection
and needs a lock on
> LogicalConnection and a deadlock occurs.

That definitely seems like a bad design; your description is quite clear
and makes the problem really stand out. Thanks for all the hard work
on this problem!

I believe that the surest means to avoid such problems is to establish
and keep to a single well-defined order of synchronization.

My intuition is that the proper order should be logical connection first,
physical connection second; do you have a sense for whether there are any
other places where we try to move in the other direction?

It is often useful to catalog the current synchronization behaviors;
sometimes I simply insert some debugging code into the Derby libraries
to capture those behaviors and run the code to observe what is occurring.
However we do it, having a nice table of code paths where we lock these
objects, and the order in which we lock them, would help us determine
what order is followed by the current code in most cases.

Once we are clear on what the correct order of synchronization should be,
there are essentially two techniques for repairing code which is violating
the ordering (i.e., code which tries to lock the primary object while holding
the lock on the secondary object):

1) Change some method higher in the call stack so that it locks the primary
object first, before calling this problematic method on the secondary object.

2) Change the problematic method itself, so it doesn't try to lock the primary
object.

So, e.g., if we determine that the problem is where ClientPooledConnection.close()
calls LogicalConnection.nullPhysicalConnection, then we could either:
1) Change the caller of ClientPooledConnection.close() to first lock the
LogicalConnection before calling close(), or
2) Change ClientPooledConnection.close() so that it doesn't call nullPhysicalConnection()

Can you provide the complete stack traces of the problematic deadlock when
it occurs?

thanks,

bryan

Mime
View raw message