[openib-general] [PATCH 3/7] IB/ipoib - Use the new verbs DMA mapping functions

[Michael S. Tsirkin]

Quoting r. Roland Dreier <rdreier at cisco.com>:
> Subject: Re: [PATCH 3/7] IB/ipoib - Use the new verbs DMA mapping functions
> 
>  > For example, ib_send_wr/ib_recv_wr could have an *optional* "void *data"
>  > field.  And we could have a rule that ULP must *either* pass in the void
>  > *data, and do dma mappings through the usual dma API, or go through the
>  > ib_dma mappings (or both).
> 
> This is a good direction to look at, because it makes "inline data"
> more useful.  However I think it just shifts the cost elsewhere,
> because now low-level drivers need to have a conditional test for
> every gather/scatter entry to see which case is being used.

I agree there's same cost to the conditional, but at least all
drivers get some benefit in return as opposed to getting
penalized for ipath limitations with nothing in return.

It might be enough to make it per-wr, not per-s/g entry - ULPs
that use s/g likely are not using low memory.

> So I'm not sure it's really a net win.

Well, this is not different from testing the "inline" flag.
So for very short messages I have tests (in user level) to show
this is a net win with respect to latency.

-- 
MST