[ofa-general] Re: [PATCH] core/cm: improve request message interpretation of subnet local fields
Jason Gunthorpe
jgunthorpe at obsidianresearch.com
Mon Sep 17 11:30:45 PDT 2007
On Mon, Sep 17, 2007 at 11:01:47AM -0700, Sean Hefty wrote:
> >The problem arises when the active Solaris client is sending a
> >connection request to a passive OFED server instance. Solaris will set
> >the hop_limit field to 0xFF and will not expect or enable GRH routing.
> >The subsequent exchange of RC messages are therefore silently dropped
> >since one side expects GRH traffic and the other doesn't.
>
> Is this an issue with the SM setting the hop_limit to 0xff or the active
> CM? Currently the ib_cm sets the local_subnet value to 1 on the active
> side. I have a patch that sets it based on the hop_limit in the path
> record. I'm trying to determine if a more complicated solution will be
> needed for ib router support. (Those changes can be separate if needed.)
I'm with Hal on this - why does this cause a problem? There is no IB
packet verification check that tests if a GRH is present, only if it
is presen it must be valid - so how can an extra correctly filled in
GRH cause anything but degraded performance?
So, it either must be that Solaris is not configuring the card to
validate the GRH properly, or the GRH fields produced by Linux are
incorrect.
I'm generally leary about overriding GRH insertion outside of the
control of the SM. I'd much prefer it if the clients never tested the
prefix to select if a GRH is needed or not. There may be useful HA
situations where a SM could route traffic for an apparently on-link
GID to a router port - same use cases as proxy arp in IP land.
Jason
More information about the general
mailing list