[ofw] Re: [ewg] RE: SC'09 BOF - Meeting notes and Final Slides

[Roland Dreier]

 > Why should clients have to emulate asynchronous behavior when the
 > (Mellanox) HCA command interface is already asynchronous?

That's one HCA command model... others may not be (eg I believe ehca
makes synchronous hypercalls to do things like modify QP, and ipath is
pure software with no other asynchronous entity to talk to).

 > I would prefer any threading be added to the HCA driver in the
 > kernel, and not as threads in each process.  That would enable a
 > seamless transition to a better driver model in the future.

Of course Windows can (and should!) design the best API for Windows;
however for Linux I would probably argue in favor of conservatism
(follow the model we've had for years when it's not totally broken) and
simplicity (it's hard for developers to get callback-based async APIs
right, and I'd rather not have an API that tries to allow both async and
sync ways of doing the same thing).

 - R.