[ofa-general] Non-Signaled RDMA Read (cxgb3)

[Caitlin Bestler]

James Lentini wrote:

> > Well if applications should handle infrequent spurious notifcations
> > where no cqe is available, then I can indeed add code in my provider
> > poll to silently throw away unsignaled read completions.
> 
> Did you consider returning an error from ib_post_send to indicate
> that an unsignalled RDMA read is not supported?
> 
> james

Suppressing notifications is a performance optimization, not a
correctness
issue. That is, while it is certainly desirable for a device to provide
no more notifications than required, the application ultimately has to
be
able to deal with spurious notifications.

Implementations ultimately have to choose between very costly
synchronization,
too few notifications or too many notifications. Clearly, slightly too
many
is far preferable than slightly too few. And how much effort is made to
keep
that gap small is a question best left to the marketplace.

Given that, I don't see why the inability to honor an optimization
request
should be flagged as an "error".