[rds-devel] [PATCH] RDS/IB: Fix memory leak on the HCA driver unload.

[Vladimir Sokolovsky]

Or Gerlitz wrote:
> Vladimir Sokolovsky wrote:
>> While removing HCA driver ib_dealloc_pd can be invoked before all 
>> existing
>> connection (QPs) closed and then it will fail on reference counter.
>>
>> Added list of connections per HCA device. All existing connections
>> will be shutted down on device removal before deallocating the PD.
> 
> Looking on the code, rds does get per connection notification on device 
> removal through the rdma-cm DEVICE_REMOVAL event and it calls 
> rds_conn_drop(), etc.
> 
> So it seems to me that the source of the problem here is using two 
> mechanisms for being notified on device removal, first through the 
> rdma-cm and second through the ib_client_register API.
> 

Yes,  you are right.

> My thinking is that the correct way to follow here, is to remove from 
> rds the code that uses the client register API and rely on the rdma-cm 
> only.
> 
> comments?
> 
 > Or.
 >

What about all per device structures: mr, fmr, pd (rds_ib_add_one)?

Vladimir.