Re: [PATCH for-next v12 6/6] RDMA/rxe: Convert mca read locking to RCU

[Jason Gunthorpe <jgg@nvidia.com>]

On Wed, Feb 23, 2022 at 04:40:19PM -0600, Bob Pearson wrote:

> OK I looked at this again. Here is what happens.
> 
> without the complete()/wait_for_completion() fix the test passes but
> the (private i.e. not visible to rdma-core) mca object is holding a
> reference on qp which will be freed after the call_rcu() subroutine
> takes place.  Meanwhile the ib_detach_mcast() call returns to
> rdma-core and since there is nothing left to do the test is over and
> the python code closes the device. But this generates a warning
> (rdma_core.c line 940) since not all the user objects (i.e. the qp)
> were freed. In other contexts we are planning on putting
> complete()/wait...()'s in all the other verbs calls which ref count
> objects for the same reason. I think it applies here too.

This explanation makes sense..

The mistake is on the QP side where it has a refcount but doesn't wait
for it to drain I suppose.

Jason