[openib-general] Re: Gen2 User verbs usage

[Michael S. Tsirkin]

Quoting r. Roland Dreier <roland at topspin.com>:
> Subject: Re: Gen2 User verbs usage
> 
>     Michael> There currently doesnt seem to exist a way for userspace
>     Michael> to know when is setting the INLINE flag possible.
> 
> I seem to remember some discussion where we said the inline flag was
> just a hint for the low-level driver, so it's always possible to set it.

I'm not against this approach, on principle, what bothers me latency vs cpu
utilization tradeoff is involved so the low-level driver may not be
the right place to take that decision.

There's also the point that for inline you dont need to pass in
a valid rkey, so the app may be better off knowing about it.

Finally, if its just a hint a separate pass over the s/g list
would be needed to calculate the size and check it fits inline,
which kind of implies performance penalty.

What is your opinion?

>     Michael> It would seem what we need is another attribute passed to
>     Michael> create_qp that would specify the max inline buffer size,
>     Michael> and probably an hca attribute to give the maximum legal
>     Michael> value for this attribute.
> 
> Do we care enough about this special feature to do this?  It seems
> Mellanox VAPI worked out OK without that.
> 
>  - R.
> 

It may be users are only recently waking to this feature.
Latency benefits for small to medium sized messages are significant.

-- 
MST - Michael S. Tsirkin