[openib-general] Re: dev_remove in the CMA

[Sean Hefty]

Tom Tucker wrote:
> I'm don't understand the dev_remove usage in the rdma_cm_id. It looks to
> me like if the user calls rdma_resolve_addr, but never calls
> rdma_resolve_route that the device could not be removed. Is this the
> intended behavior? 

Once a rdma_cm_id has been bound to a device, the user must destroy that 
rdma_cm_id on device removal before the device removal can proceed.

> Is the goal to prevent the user from removing the device if the client
> is in a callback? If so, can't we just increment and decrement in the
> cma_notify_user function?  I guess I just don't understand...

Users above the CMA obtain their device pointer from the CMA.  The device 
pointer must be valid outside of CMA callbacks, so device removal is delayed 
until the user releases all resources associated with a device.  Destruction of 
the rdma_cm_id indicates that the user is no longer using that device.

- Sean