[ofa-general] [PATCH] osm: bug in fat-tree routing

[Yevgeny Kliteynik]

Hi Hal,

Fixing bug in fat-tree routing with loops in fabric.

When the switch connections form loop in the fabric, fat-tree routing
could follow this loop and set LFT table value for switches that have
been already configured for this LID.

Please apply to ofed_1_2 and to master.

Thanks.

-- Yevgeny

Signed-off-by:  Yevgeny Kliteynik <kliteyn at dev.mellanox.co.il>
---
 osm/opensm/osm_ucast_ftree.c |   17 +++++++++++++++++
 1 files changed, 17 insertions(+), 0 deletions(-)

diff --git a/osm/opensm/osm_ucast_ftree.c b/osm/opensm/osm_ucast_ftree.c
index a4f307d..655a821 100644
--- a/osm/opensm/osm_ucast_ftree.c
+++ b/osm/opensm/osm_ucast_ftree.c
@@ -1826,6 +1826,23 @@ __osm_ftree_fabric_route_upgoing_by_goin
          set LFT(target_lid) on the remote switch to the remote port */
       p_remote_sw = p_group->remote_hca_or_sw.remote_sw;
 
+      if ( osm_switch_get_least_hops(p_remote_sw->p_osm_sw, 
+                                     cl_ntoh16(target_lid)) != OSM_NO_PATH )
+      {
+         /* Loop in the fabric - we already routed the remote switch 
+            on our way UP, and now we see it again on our way DOWN */
+         osm_log(&p_ftree->p_osm->log, OSM_LOG_DEBUG,
+                 "__osm_ftree_fabric_route_upgoing_by_going_down: "
+                 "Loop of lenght %d in the fabric:\n                             "
+                 "Switch %s (LID 0x%x) closes loop through switch %s (LID 0x%x)\n",
+                 (p_remote_sw->rank - highest_rank_in_route) * 2,
+                 __osm_ftree_tuple_to_str(p_remote_sw->tuple),
+                 cl_ntoh16(p_group->base_lid),
+                 __osm_ftree_tuple_to_str(p_sw->tuple),
+                 cl_ntoh16(p_group->remote_base_lid));
+         continue;
+      }
+
       /* Four possible cases:
        *
        *  1. is_real_lid == TRUE && is_main_path == TRUE: 
-- 
1.4.4.1.GIT