Hi Hal,
On 13:27 Wed 25 Aug , Hal Rosenstock wrote:
>
> I'm seeing an issue with ibnetdiscover from a CA port where it appears
> to extend a path at a "remote" CA port (it's actually another port on
> the same CA) to query NodeInfo of the next hop beyond it. I get the
> following error message:
>
> src/query_smp.c:188; umad (DR path slid 0; dlid 0; 0,1,20,2 Attr
> 0x11:0) bad status 110; Connection timed out
>
> where smpquery -D nodeinfo of 0,1,20 is a CA which can also be seen
> from the topology.
>
> It appears to stem from the following code snippet from
> libibnetdisc/src/ibnetdisc.c:recv_port_info
>
> if (port_num && mad_get_field(port->info, 0, IB_PORT_PHYS_STATE_F)
> == IB_PORT_PHYS_STATE_LINKUP
> && ((node->type == IB_NODE_SWITCH && port_num != local_port) ||
> (node == fabric->from_node && port_num == local_port))) {
> ib_portid_t path = smp->path;
> if (extend_dpath(engine, &path, port_num) > 0)
> query_node_info(engine, &path, node);
> }
This makes sense for me.
>
> that was introduced by:
> commit fcb8d5e7588e38508a8e354c37009d73c0a3889f
> Author: Sasha Khapyorsky <[email protected]>
> Date: Sat Apr 10 02:43:24 2010 +0300
>
> libibnetdisc: no backward NodeInfo queries
>
> Then switch is reached via port N we don't need to query back via this
> port - source node is discovered already. Finally this saves some amount
> of unnecessary MADs.
>
> Signed-off-by: Sasha Khapyorsky <[email protected]>
>
> and subsequently modified by:
> commit 49d149c63a44d99259f516a15af53d8cf3f0e7c9
> Author: Sasha Khapyorsky <[email protected]>
> Date: Tue Apr 13 19:54:45 2010 +0300
>
> libibnetdisc: don't try to cross discovery over CA
>
> When discovery is running from CA node it shouldn't try to cross over
> all ports, but only via local one (send over non-local ports will fail
> since CA doesn't route MADs).
>
> Signed-off-by: Sasha Khapyorsky <[email protected]>
>
> due to the (node == fabric->from_node && port_num == local_port)
> clause being TRUE.
But I don't see how those patches are actually related to the story. An
original (before patches) condition was:
if (port_num && mad_get_field(port->info, 0, IB_PORT_PHYS_STATE_F)
== IB_PORT_PHYS_STATE_LINKUP
&& (node->type == IB_NODE_SWITCH || node == fabric->from_node))
, which has the described bug as I can understand this.
Sasha
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
