Trunk warnings: "...Handles still allocated"

[jayesh]

When I run pt2pt/scancel & some io tests on trunk I get the following message when compiled with "--enable-strict --enable-g=all",

shakey:/sandbox/jayesh/freshBuild/mpich2/test/mpi/pt2pt> mpiexec -n 2 scancel
 No Errors
In direct memory block for handle type REQUEST, 3 handles are still allocated

Regards, Jayesh

[gropp]

This is indicating that there is a bug in the handling of cancel of send requests - the requests aren't properly being freed. I checked the test code, and the test code is not at fault.

[balaji]

Milestone mpich2-1.1.2 deleted

[balaji]

I believe Dave is looking into these errors.

[goodell]

I've got a bead on this, although the proper fix might be more intrusive and riskier than we are willing to tolerate for a bug-fix release like 1.2.1. I'll update this ticket when I know more.

In general, this investigation makes it very clear that the rules for reference counting requests are not obvious or documented anywhere easy to find. We need to remedy this by formulating a coherent strategy for request refcounting, documenting it in the wiki and in code comments, and auditing the current code to make sure that we are following this. Requests are the most commonly created/destroyed object in MPICH2 and properly managing their lifecycle is essential to avoid serious memory leaks.

[goodell]

Moving to 1.3. The fix for this case will almost certainly have some subtle bug in it that we don't want to introduce for a minor release like this.