[openib-general] Re: openIB gen2 user space verbs API

[Michael S. Tsirkin]

Quoting r. Roland Dreier <roland at topspin.com>:
> Subject: Re: openIB gen2 user space verbs API
> 
>     Michael> An application would need to know what values is it legal
>     Michael> to pass to create_qp.  Maybe it makes sence to implement
>     Michael> something like query_hca, and let it return the maximum
>     Michael> legal value?
> 
> The Mellanox VAPI just used the maximum number of sg entries for the
> send queue to calculate the inline data value for a QP.  Does it make
> sense to change this interface?
> 
>  - R.
> 

It was actually you who first proposed the change :) But I'd like
to defend that decision.

I think applications have different latency/CPU utilization tradeoffs.
A microbenchmark may want to push as much data as possible inline,
another application may not.

So I think what you end up doing with VAPI API is applications starting with
the inline size they actually want and hardcoding the tavor size to s/g entries
ratio to get the right values in query qp.

If they do it right they will even work with other HCAs, just more
slowly (say the app checks qp properties and sees it didnt get the inline size
that it wanted, so it doesnt use inline), but isnt what I proposed cleaner?

Since its an ABI change anyway - lets do it right?

Its easy enough to implement - just say the word ...

-- 
MST - Michael S. Tsirkin