[ofa-general] Re: IPOIB CM (NOSRQ)[PATCH V3] patch for review

[Michael S. Tsirkin]

> > > Note 1: I have retained the code to avoid IB_WC_RETRY_EXC_ERR while
> > > performing interoperability tests As discussed in this mailing list that
> > > may be a CM bug or have the various HCA address it. Hence I would like to
> > > seperate out that issue from this patch.  At a future point when the issue
> > > gets resolved I can provide another patch to change the retry_count values
> > > back to 0 if need be.
> > 
> > The correct way to separate it, in my opinion, is to set retry_count = 0,
> > and (for now) apply a work-around patch at your site before testing.
> > We really don't want to paper over this bug, in my opinion.
> 
> Ok, will reset this back to 0, but that is not (my) preferred way. If some
> one were to pick up the code and try it with retry_count=0, the HCAs will 
> not inter-operate as is. Hence the hesitation.

BTW, why do you ignore the option to use UC QP?
Even taking this single issue aside, I think that 
UC is a better QP type choice for IPoIB than RC - you get away from RNR errors
(so you can prepost less data, and you can even reset some RQs
 temporarily, moving WRs between them, without affecting TX),
and you get send completion sooner,
so you can use less memory for send buffers and smaller TX queues.

With UC, we might get stale TX connections, so a way to detect
and handle them will need to be designed.

-- 
MST