[openib-general] tavor quirks etc (opensm compliance etc)

[Michael S. Tsirkin]

> 3rd Eitan/Michael: what is the bigger picture here? what is the 
> dependency between these four patches

In short, [2] is an independent fix to improve tavor performance.
Other things are not directly related. Detail below.

> +1 osm:Fix PathRecord bug MTU/rate/PktLife explicitly ignoring selectors
> +2 osm: tavor quirk
> +3 IB/rdmacm: tavor quirk
> +4 IB/ipoib: use appropriate mtu selector for path queries

In the above:
[1] is a bug fix I think. It is not required for [2].
[2] is a feature that improves performance for tavor without need for
    any other stack/ULP changes
[3] is a hack that should have same effect as [2] for old SMs, but it needs
    manual tuning by user. If activated, it unfortunately triggers a bug in opensm
    that [1] fixes. So it might not be a good idea after all.
[4] is not strictly necessary, and not related to this patch set -
    it just happens to also play with MTU selector.
    It is a strict compliance cleanup that I just happened to notice when
    I invented [2].

> for example is it correct that:
> 
> if [2] is applied on the SA side then [4] must be applied on ipoib else 
> if will get 1K mtu on its path query?

Not really - ipoib does not actually use the MTU it gets from the path query,
according to spec it uses the bcast group mtu for all packets.

> if [2] is not applied on the SA side, then [3] is useless?

No. If [2] is applied on te SA side, the [3] is unnecessary.


-- 
MST