[openib-general] Re: [PATCH] change Mellanox SDP workaround to a moduleparameter

[Ralph Campbell]

I'm not all that familiar with the MTU negotiation protocol
either.  Note that the CM or SDP should be using ib_query_port()
to get the max MTU the device/port supports.

My main concern is the device specific code which is compiled
into sdp_cm_path_complete() to reduce the MTU to something
less than the maximum to work around a performance bug.
I was proposing to make this a module parameter but perhaps
a better solution would be some interface where the device
driver can set it or SDP queries the device for the optimum
value.

On Wed, 2006-02-15 at 23:24 +0200, Michael S. Tsirkin wrote:
> Quoting r. Ralph Campbell <ralphc at pathscale.com>:
> > Subject: [PATCH] change Mellanox SDP workaround to a moduleparameter
> > 
> > This patch changes the hardwired MTU limit of 1024 in SDP
> > into a module parameter so it can be disabled for HCAs
> > without the RC performance problem.
> > 
> > Signed-off-by: Ralph Campbell <ralphc at pathscale.com>
> 
> I thought about it some more: what happens if nodes with different
> max MTU values try to connect?
> 
> My understanding of the spec:
> 
> - passive side should send REJ indicating error and passing the
>   maximum MTU it supports
> - active side should retry the connection with a lower MTU
> 
> Right?
> By the way, this handling seems to be missing in CMA.
> 
-- 
Ralph Campbell <ralphc at pathscale.com>