[openib-general] [RFC] userspace IB SA support

Hal Rosenstock halr at voltaire.com
Fri Jan 12 10:33:55 PST 2007


On Fri, 2007-01-12 at 12:29, Sean Hefty wrote:
> >>Would we be okay with extending the IOCTL interface to allow multicast joins, 
> >>notice registration, and event reporting?  Or would it be acceptable to change 
> >>the ib_umad read/write interface to add a command?
> > 
> > 
> > What do you have in mind here ?
> 
> I was thinking of one of two possibilities here.  Currently there are IOCTL 
> calls to register/unregister with the MAD layer.  Additional IOCTL calls could 
> be added to join/leave multicast groups and register/unregister for SA events. 
> Multicast and SA events would need to be reported through another IOCTL of some 
> sort.
> 
> The alternative basically rewrites the ib_umad interface to allow read and write 
> to carry some sort of command, rather than mapping them directly to sending and 
> receiving a MAD.  This is how most of the RDMA kernel to user interfaces are 
> written.  For example, let read return an event type (MAD received, multicast 
> event, etc.), along with the event data (the MAD, etc.).

Do we really want to go down this approach ?

> >>>As an alternative, a new kernel userspace SA module could be created to
> >>>explicitly interface with the kernel ib_sa.
> > 
> > IMO, this is the best way to go.
> 
> This was my original approach a couple of months back, but wasn't accepted as 
> mer gable upstream because it increased the size of the user to kernel 
> interface. 

Can you point me at this ? I must have missed it.

-- Hal

>  If we can agree that this approach is usable, we can discuss more 
> specific implementation details.
> 
> - Sean





More information about the general mailing list