[openib-general] Re: [PATCH] Opensm - client reregistration bit

[Hal Rosenstock]

On Wed, 2005-11-23 at 04:37, Yael Kalka wrote:
> Hi Hal,
> 
> Currently when sending mads with PortInfo Set we set client_rereg to
> be zero

I think you mean one here.

>  if we are in first_time_master_sweep and relevant capability
> mask is on). On other cases - we send in client_reregister bit the
> data we have saved. If this data is 1 - then we will keep on sending
> it.
> This patch assures that we send 0 in the client_reregister bit,
> unless we specifically want to send 1.

Thanks. Applied.

> I think there is still a bug in the client_reregisteration. If we
> merge subnets, the new ports will not be set with client_rereg=1
> (since the master SM is not in first sweep).
> I plan to continue working on the client_reregsitration issue next
> week, and fix this issue as well.

OK. Thanks.

There is also the issue that client_reregistration is set too early and
often the SA is not ready as the subscriptions are received before the
first sweep completes. So this currently relies on the client timeout
and retransmit strategy. The SM should only request this when it is
really ready.

-- Hal

> Thanks,
> Yael
> 
> Signed-off-by:  Yael Kalka <yael at mellanox.co.il>
> 
> Index: opensm/osm_lid_mgr.c
> ===================================================================
> --- opensm/osm_lid_mgr.c	(revision 4119)
> +++ opensm/osm_lid_mgr.c	(working copy)
> @@ -1146,6 +1146,8 @@ __osm_lid_mgr_set_physp_pi(
>    if ( ( p_mgr->p_subn->first_time_master_sweep == TRUE ) &&
>         ( (p_old_pi->capability_mask & IB_PORT_CAP_HAS_CLIENT_REREG) != 0 ) )
>      ib_port_info_set_client_rereg( p_pi, 1 );
> +  else
> +    ib_port_info_set_client_rereg( p_pi, 0 );
>  
>    /* We need to send the PortInfoSet request with the new sm_lid
>       in the following cases: