[ofa-general] Re: Re: [PATCH RFC] sharing userspace IB objects

Michael S. Tsirkin mst at dev.mellanox.co.il
Mon Jul 2 11:39:06 PDT 2007


> Quoting Roland Dreier <rdreier at cisco.com>:
> Subject: Re: [ofa-general] Re: Re: [PATCH RFC] sharing userspace IB objects
> 
>  > > Correct. The number of messages in flight per EEC is 1 per IB spec.
>  > > The fact that IB requires SQ WQEs to complete in order, even if their 
>  > > destination is different EECs,
>  > 
>  > Where's this requirement in the spec (and could this be relaxed as it
>  > seems like it is overly "specified") ? Just wondering...
> 
> I don't think we want to relax the requirement that work requests
> complete in order.  It's hard enough to get applications correct
> without having to worry about out-of-order completions,

Hmm, they seem to deal fine with this in case of SRQ. Why not here?

I guess this depends on the application, but let's look at
something like IPoIB or SDP: all we do when we get a send
completion is look up a WR a free it. It won't be too hard
to deal with out of order, either. If an app uses a pointer
as WRID, it's even easier.

> and I think
> specifying all the corner cases would be a nightmare.  Eg do we allow
> successful completions after a completion with error?  and so on...

However, as Dror notes, the in-order requirement simply moves
the complexity to hardware. Which might be one of the reasons why
there are no HW implementations of RD out there.

-- 
MST



More information about the general mailing list