[ofa-general] [RFC PATCH 4/4] rdma/cma: implementRDMA_ALIGN_WITH_NETDEVICE ha mode

[Sean Hefty]

>> Maybe instead of making this a cm_id event, we should add a concept of
>> rdma async events that aren't necessarily affiliated with any
>> particular cm_id?  IE a new channel for these types.  Then you can
>> post it once when a route changes that affects rdma devices, for
>> example...
>>
>As opposed to posting the event to every cm_id affected...

I thought about this, and I agree that it's worth exploring.  The locking to
support device removal ended up being fairly complex.  (I'm not sure it would
have been any easier for ULPs to do this though.)  The main counter I see to
using a separate channel is that device removal is invoked per rdma_cm_id, so
there's precedence for invoking the callback per id.

My expectation is that this is a rare event.

- Sean